A quality assessment method, system, device and medium for equipment spare parts

Abstract

The present invention relates to the field of data processing technology, and discloses a quality assessment method, system, device and medium for equipment spare parts. The method includes: obtaining a target sample data set of equipment spare parts; constructing a feature extraction model based on a preset fully connected neural network, and performing preset feature extraction processing on the target sample data set based on the feature extraction model to obtain a first feature set; updating the first network parameter of the feature extraction model based on the first feature set, and returning to the step of performing preset feature extraction processing on the target sample data set until the feature extraction model meets the first preset training condition; constructing a target assessment model based on the trained feature extraction model to perform quality assessment on the equipment spare parts sample data to be assessed. Through the scheme of the present invention, accurate equipment spare parts quality assessment results can be automatically output, reducing manual intervention and reducing costs.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a quality assessment method, system, device and medium for equipment spare parts. Background Art

[0002] With the development of automation and intelligent technology, traditional manual inspection methods can no longer meet the needs of high-efficiency and high-precision quality assessment. Especially in the spare parts manufacturing industry, how to quickly and accurately evaluate the quality of spare parts and ensure that they strictly follow industrial standards has become an important issue facing the manufacturing industry. At present, although a variety of quality assessment solutions have emerged in the market, these solutions generally face the dual dilemma of insufficient adaptability and data processing bottlenecks. On the one hand, some solutions rely too much on preset inspection standards and manual experience, and it is difficult to flexibly respond to complex changes and diverse needs in the production process. On the other hand, faced with the emergence of massive data, traditional methods are particularly struggling in terms of processing speed and data analysis capabilities. The above-mentioned related technologies are often highly dependent on manual operations, which not only pushes up costs, but also seriously restricts the further improvement of inspection efficiency and accuracy.

[0003] Therefore, it is urgent to propose a more intelligent quality assessment method, system, equipment and medium for equipment spare parts to overcome the above limitations. Summary of the invention

[0004] In view of this, the present invention proposes a quality assessment method, system, equipment and medium for equipment spare parts, which solves the problems of complicated process, high difficulty, high labor cost and low accuracy in the quality assessment process of equipment spare parts, releases the production capacity of production personnel and improves the efficiency and accuracy of quality assessment.

[0005] Based on the above purpose, an embodiment of the present invention provides a method for quality assessment of equipment spare parts, which specifically includes the following steps:

[0006] Obtain target sample data set of equipment spare parts;

[0007] Building a feature extraction model based on a preset fully connected neural network, and performing preset feature extraction processing on the target sample data set based on the feature extraction model to obtain a first feature set;

[0008] Based on the first feature set, a first network parameter of the feature extraction model is updated, and the step of performing a preset feature extraction process on the target sample data set is returned until the feature extraction model meets a first preset training condition;

[0009] Based on the trained feature extraction model, a target evaluation model is constructed to perform quality evaluation on sample data of equipment spare parts to be evaluated.

[0010] In some implementations, the step of performing preset feature extraction processing on the target sample data set based on the feature extraction model to obtain a first feature set includes:

[0011] In response to the input layer in the feature extraction model receiving the target sample data set, the target sample data set is used as target input data, and the next neural network layer of the input layer is used as the target network layer;

[0012] Performing a preset linear change process on the target input data based on the target network layer, and performing a preset feature extraction process on the target input data after the linear change based on a ReLU activation function, so as to obtain a feature result output by the target network layer;

[0013] Using the feature result as new target input data, using the next neural network layer of the target network layer as the new target network layer, and returning to the step of performing the preset linear change processing on the target input data based on the target network layer until there is no next neural network layer in the target network layer, and using the feature result as the first feature set;

[0014] The next neural network layer includes the hidden layer or the output layer.

[0015] In some implementations, the step of updating a first network parameter of the feature extraction model based on the first feature set includes:

[0016] Determining a first loss value of the output layer according to the first feature set;

[0017] Based on the first loss value, determine the first weight gradient value and the first bias gradient value corresponding to each of the hidden layers respectively;

[0018] Based on each of the first weight gradient values ​​and each of the first bias gradient values, predicting a second weight gradient value and a second bias gradient value respectively corresponding to each of the hidden layers;

[0019] Based on each of the first weight gradient values, each of the first bias gradient values, each of the second weight gradient values, and each of the second bias gradient values, the weight values ​​and bias parameters corresponding to each of the hidden layers are updated to update the first network parameters of the feature extraction model.

[0020] In some embodiments, the step of constructing a target evaluation model based on the trained feature extraction model includes:

[0021] Obtaining a target feature set corresponding to the target sample data set output by the trained feature extraction model;

[0022] Based on the target feature set, a feature dimensionality reduction model is trained, and a target dimensionality reduction set corresponding to the target feature set output by the trained feature dimensionality reduction model is obtained;

[0023] Based on the target dimension reduction set, training a classifier model;

[0024] A target evaluation model is constructed based on the trained feature extraction model, the trained feature dimension reduction model and the trained classifier model.

[0025] In some embodiments, the step of training a feature dimensionality reduction model based on the target feature set includes:

[0026] Based on the feature dimensionality reduction model, performing a preset feature dimensionality reduction process on the target feature set to generate a first feature dimensionality reduction set;

[0027] Determining a second loss value of the feature dimensionality reduction model based on the first feature dimensionality reduction set;

[0028] Based on the second loss value, the second network parameter of the feature dimensionality reduction model is updated, and the step of performing a preset feature dimensionality reduction process on the target feature set is returned until the feature dimensionality reduction model meets the second preset training condition, thereby determining the trained feature dimensionality reduction model.

[0029] In some embodiments, the step of training a classifier model based on the target dimension reduction set includes:

[0030] Based on the classifier model, each sample data in the target dimension reduction set is subjected to a preset classification process to generate a prediction data set;

[0031] Determining a third loss value of the classifier model based on all predicted data sets;

[0032] Based on the third loss value, the third network parameter of the classifier model is updated, and the step of performing preset classification processing on each sample data in the target dimensionality reduction set is returned until the classifier model meets the third preset training condition, thereby determining the trained classifier model.

[0033] In some implementations, the step of obtaining a target sample data set of equipment spare parts includes:

[0034] Acquire an original sample data set of the equipment spare parts, and mark each sample data in the original sample data set, wherein the marked categories include qualified and unqualified;

[0035] Based on the data expansion model, performing data expansion processing on the annotated original sample data set to generate a first expanded data set;

[0036] Obtaining a true probability corresponding to the first expanded data set, updating a fourth network parameter of the data expansion model based on the true probability, and returning to the step of performing data expansion processing on the labeled original sample data set until the data expansion model meets a fourth preset training condition, and determining the trained data expansion model;

[0037] Based on the trained data expansion model, data expansion processing is performed on the labeled original sample data set to generate a second expanded data set, and based on the second expanded data set, the target sample data set is obtained.

[0038] Another aspect of the embodiment of the present invention further provides a quality assessment system for equipment spare parts, including:

[0039] An acquisition unit configured to acquire a target sample data set of equipment spare parts;

[0040] A feature extraction unit is configured to construct a feature extraction model based on a preset fully connected neural network, and perform preset feature extraction processing on the target sample data set based on the feature extraction model to obtain a first feature set;

[0041] A training unit configured to update a first network parameter of the feature extraction model based on the first feature set, and return to the feature extraction unit to perform a preset feature extraction process on the target sample data set again until the feature extraction model meets a first preset training condition;

[0042] The evaluation unit is configured to construct a target evaluation model based on the trained feature extraction model to perform quality evaluation on the sample data of equipment spare parts to be evaluated.

[0043] According to another aspect of an embodiment of the present invention, a computer device is provided, comprising: at least one processor; and a memory, wherein the memory stores a computer program executable on the processor, and the computer program implements the steps of the above method when executed by the processor.

[0044] According to another aspect of the embodiments of the present invention, a computer-readable storage medium is provided, which stores a computer program that implements the above method steps when executed by a processor.

[0045] The present invention has at least the following beneficial technical effects: the quality assessment method for equipment spare parts of the present invention is based on a 6-layer fully connected neural network structure to encapsulate a feature extraction model, and learns and extracts deep, nonlinear features in the data through the output layer and multiple hidden layers, which helps to capture subtle differences and complex patterns that are difficult to find with traditional methods, thereby improving the accuracy and reliability of subsequent quality assessments. An exploratory gradient descent method is used to update the first network parameters of the feature extraction model until the preset training conditions are met, so that the feature extraction model can continuously learn and improve from the data, avoiding the problems of gradient vanishing, gradient explosion or falling into a local optimal solution, and ultimately obtaining a more accurate and generalized feature representation, which helps to improve the robustness of the assessment model, and constructs a target assessment model based on the trained feature extraction model to perform quality assessment of equipment spare parts sample data, thereby reducing the cost of manual operations, and having multiple beneficial effects such as high efficiency, accuracy, reliability, strong generalization ability, high degree of automation, and support for rapid response. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these drawings without paying creative work.

[0047] Figure 1 A block diagram of an embodiment of a method for evaluating the quality of equipment spare parts provided by the present invention;

[0048] Figure 2 A schematic diagram of an embodiment of a feature extraction model training process provided by the present invention;

[0049] Figure 3 A schematic diagram of an embodiment of a feature dimensionality reduction model training process provided by the present invention;

[0050] Figure 4 A schematic diagram of an embodiment of a quality assessment system for equipment spare parts provided by the present invention;

[0051] Figure 5 A schematic diagram of the structure of an embodiment of a computer device provided by the present invention;

[0052] Figure 6 A schematic diagram of the structure of an embodiment of a computer-readable storage medium provided by the present invention. DETAILED DESCRIPTION

[0053] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the embodiments of the present invention are further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0054] It should be noted that all expressions using "first" and "second" in the embodiments of the present invention are for distinguishing two non-identical entities with the same name or non-identical parameters. It can be seen that "first" and "second" are only for the convenience of expression and should not be understood as limitations on the embodiments of the present invention. The subsequent embodiments will not explain this one by one.

[0055] Based on the above purpose, the first aspect of the embodiment of the present invention provides an embodiment of a quality assessment method for equipment spare parts. Figure 1 As shown, it includes the following steps:

[0056] Step S100, obtaining a target sample data set of equipment spare parts;

[0057] Step S200, constructing a feature extraction model based on a preset fully connected neural network, and performing preset feature extraction processing on the target sample data set based on the feature extraction model to obtain a first feature set;

[0058] Step S300, updating a first network parameter of a feature extraction model based on a first feature set, and returning to the step of performing a preset feature extraction process on a target sample data set until the feature extraction model meets a first preset training condition;

[0059] Step S400: construct a target evaluation model based on the trained feature extraction model to perform quality evaluation on sample data of equipment spare parts to be evaluated.

[0060] In some implementations, the target sample data set is a sample data set obtained after data expansion. The preset fully connected neural network is a 6-layer fully connected neural network, and the structure of the 6-layer fully connected neural network is as follows:

[0061] a-Input layer: receives and transmits sample data in the target sample data set;

[0062] b-Hidden layer: Each layer consists of a certain number of neurons;

[0063] b1-first hidden layer: has 128 neurons, used for preliminary feature extraction;

[0064] b2-second hidden layer: has 256 neurons to further abstract features;

[0065] b3-the third hidden layer: has 512 neurons, further deepening feature extraction;

[0066] b4-the fourth hidden layer: has 1024 neurons, further deepening feature extraction;

[0067] b5 - output layer: has 512 neurons.

[0068] Through the gradual deepening of four hidden layers (b1 to b4), the network is able to learn multi-level features from simple to complex. The first hidden layer (b1) initially extracts basic features, and the subsequent hidden layers further abstract and deepen these features until the fourth hidden layer (b4) reaches a higher level of feature complexity and abstraction. This multi-level structure helps capture richer and more subtle information in the data. As the number of layers increases and the number of neurons increases, the dimension and complexity of the features that the network can learn also increase, thereby improving the efficiency and quality of feature representation. The output layer (b5) is responsible for aggregating and integrating the features extracted by all previous hidden layers. The output layer with 512 neurons can accommodate a large amount of feature information, and through the interaction and combination of these neurons, a more comprehensive and in-depth understanding of the input data is formed.

[0069] In some embodiments, in view of the challenges often encountered by neural networks in the feature extraction process in related technologies, such as gradient disappearance, gradient explosion, and easy to fall into local optimal solutions, these problems directly affect the stability of feature extraction model training and the final performance. In order to solve this pain point, on the basis of the above-mentioned 6-layer fully connected neural network, an exploratory gradient descent strategy is further introduced to improve the feature extraction model. The exploratory gradient descent strategy introduces a forward exploration mechanism based on the current gradient information of the feature extraction model. Specifically, before each gradient update, not only the current gradient state is analyzed, but also the possible change trajectory of the gradient in the future is predicted prospectively. Based on this prediction information, the algorithm can intelligently adjust the learning step size and direction of the feature extraction model, thereby realizing dynamic optimization in the gradient descent process. This exploratory method significantly improves the efficiency and effect of the training process, not only accelerates the convergence speed of the model, but also helps the feature extraction model escape the trap of local optimal solution.

[0070] In some embodiments, Figure 2As shown, the target sample data set obtained after expansion is input into the initialized feature extraction model to train the feature extraction model. The sample data in the target sample data set is subjected to linear change processing and nonlinear activation processing through each hidden layer in the feature extraction model, and the feature representation corresponding to the sample data in the target sample data set is further calculated by the output layer, i.e., the first feature set. On the basis of the first feature set, the weights and bias parameters (i.e., the first network parameters) of each hidden layer and output layer in the feature extraction model are updated by an exploratory gradient descent strategy. It is continuously iterated until the preset stop condition (i.e., the first preset training condition) is met to complete the training of the feature extraction model. The preset stop condition is to reach the preset maximum number of iterations, for example, 1000 times. The combination of a 6-layer fully connected neural network with an exploratory gradient descent strategy not only enhances the stability of the feature extraction model, but also greatly improves the accuracy and generalization ability of the feature extraction model, laying a solid foundation for subsequent model applications.

[0071] The quality assessment method for equipment spare parts of the present invention is based on a 6-layer fully connected neural network structure to encapsulate a feature extraction model, and learns and extracts deep, nonlinear features in the data through the output layer and multiple hidden layers, which helps to capture subtle differences and complex patterns that are difficult to find with traditional methods, thereby improving the accuracy and reliability of subsequent quality assessments. An exploratory gradient descent method is used to update the first network parameters of the feature extraction model until preset training conditions are met, so that the feature extraction model can continuously learn and improve from the data, avoiding the problems of gradient vanishing, gradient explosion or falling into a local optimal solution, and ultimately obtaining a more accurate and generalized feature representation, which helps to improve the robustness of the assessment model. A target assessment model is constructed based on the trained feature extraction model to perform quality assessment of equipment spare parts sample data, thereby reducing the cost of manual operation, and having multiple beneficial effects such as high efficiency, accuracy, reliability, strong generalization ability, high degree of automation, and support for rapid response.

[0072] In some embodiments, based on the feature extraction model, a preset feature extraction process is performed on the target sample data set to obtain a first feature set, including: in response to the input layer in the feature extraction model receiving the target sample data set, the target sample data set is used as the target input data, and the next neural network layer of the input layer is used as the target network layer; based on the target network layer, a preset linear change process is performed on the target input data, and based on the ReLU activation function, a preset feature extraction process is performed on the linearly changed target input data to obtain a feature result output by the target network layer; the feature result is used as the new target input data, and the next neural network layer of the target network layer is used as the new target network layer, and the step of performing a preset linear change process on the target input data based on the target network layer is returned until there is no next neural network layer in the target network layer, and the feature result is used as the first feature set, wherein the next neural network layer includes a hidden layer or an output layer.

[0073] In some embodiments, the step of updating the first network parameters of the feature extraction model based on the first feature set includes: determining the first loss value of the output layer according to the first feature set; determining the first weight gradient value and the first bias gradient value corresponding to each hidden layer based on the first loss value; predicting the second weight gradient value and the second bias gradient value corresponding to each hidden layer based on each first weight gradient value and each first bias gradient value; updating the weight parameters and bias parameters corresponding to each hidden layer based on each first weight gradient value, each first bias gradient value, each second weight gradient value, and each second bias gradient value, so as to update the first network parameters of the feature extraction model.

[0074] In some implementations, the training process of the feature extraction model is described:

[0075] Step 1: Initialize the network parameters of the neural network in the feature extraction model, that is, the first network parameters, including weight parameters and bias parameters. The initialization method can be expressed as:

[0076] ;

[0077] ;

[0078] In the formula, is the weight matrix of the i-th layer of the neural network; is the bias vector of the i-th layer of the neural network; is the scaling factor of the neural network weights, is the scaling factor of the bias of the neural network; represents a normal distribution with a mean of 0 and a standard deviation of 1; is a tiny noise. Preferably, is a noise vector whose elements are all 0.001.

[0079] Step 2: Input the target sample data set through each layer of the neural network in the feature extraction model. Each layer performs linear transformation and nonlinear activation function processing on the input data, and finally obtains the feature representation of the output layer, that is, the feature result. Except for the next hidden layer of the input layer, each layer of the neural network uses the feature result of the previous layer as input data, and the feature result of the output of each layer of the neural network can be expressed as:

[0080] ;

[0081] In the formula, Represents the neural network The activation output of the layer, i.e. The characteristic results of the layer; The neural network The output of the layer, i.e. The characteristic results of the layer; is the adaptive ReLU activation function. Indicates preset linear change processing.

[0082] Assume that the input of the adaptive ReLU activation function is , the calculation method of the adaptive ReLU activation function can be expressed as:

[0083] ;

[0084] In the formula, and are the adaptive coefficients of the positive and negative parts respectively, To obtain the maximum value function, is the minimum value function. Preferably, Set to 0.3, Set to 0.7. The above Rez( ) represents the nonlinear activation function processing, that is, the preset feature extraction processing.

[0085] Through this step, the characteristic results of each layer output of the neural network can be obtained, and the characteristic results finally output by the output layer are determined as the first feature set, so as to facilitate the subsequent updating of the network parameters of each layer of the neural network.

[0086] Step 3: Adopt an exploratory gradient descent update strategy. Before each update of the neural network parameters, predict the possible gradient changes. Adjust the step size and direction of the gradient update according to the prediction results, optimize the parameter adjustment during the training process, and avoid falling into the local optimum too early. The calculation method is expressed as:

[0087] ;

[0088] ;

[0089] In the formula, The neural network The update amount of the layer weight parameters, The neural network is The update amount of the layer bias parameters; is the learning rate of the neural network; is the loss function with respect to the parameter The gradient of is the loss function with respect to the parameter The gradient of is the loss function of the neural network; is the exploration intensity coefficient; is the prediction function of the gradient. Preferably, Set to 0.03, Cross entropy loss is used.

[0090] The exploratory gradient prediction function uses historical gradient data for linear regression prediction. The calculation method is expressed as:

[0091] ;

[0092] In the formula, Indicated in The loss function of the iteration is about the parameter The gradient of is the attenuation factor; To explore the strength coefficient; is the total number of iterations. Preferably, Set to 0.98, Set to 2.

[0093] Furthermore, the weight parameters and bias parameters of the neural network are updated by using the error back propagation method. During the training process, the neurons in the output layer receive the gradient information from the loss function and pass this gradient information to the previous hidden layer through the back propagation algorithm. By continuously adjusting the first network parameters through the optimization algorithm (such as the gradient descent method), the characteristic results of the output layer can be made closer to the target value, thereby improving the performance of the entire network. Loss function It is calculated based on the first feature set output by the output layer.

[0094] Step 4: Repeat the above steps until the first preset training condition is met, which means the model training is completed. For example, the number of iterations reaches 1000.

[0095] The quality assessment method for equipment spare parts of the present invention adopts an exploratory gradient descent neural network in the feature extraction model, optimizes the gradient update strategy, solves the problems of gradient vanishing, gradient explosion or falling into local optimal solutions that may be encountered in traditional neural network training, and improves the generalization ability of features.

[0096] In some embodiments, the step of constructing a target evaluation model based on the trained feature extraction model includes: obtaining a target feature set corresponding to a target sample data set output by the trained feature extraction model; training a feature dimension reduction model based on the target feature set, and obtaining a target dimension reduction set corresponding to the target feature set output by the trained feature dimension reduction model; training a classifier model based on the target dimension reduction set; and constructing a target evaluation model based on the trained feature extraction model, the trained feature dimension reduction model, and the trained classifier model.

[0097] In some embodiments, a target evaluation model is obtained by combining a trained feature extraction model, a data dimension reduction model, and a classifier model to perform quality evaluation on new spare parts samples and output quality evaluation results. The original sample data set of the collected new spare parts samples is input into the target evaluation model, and feature processing is performed in turn through the trained feature extraction model and the trained feature dimension reduction model. Further, the processed features are input into the trained classifier model for classification, and then the classification result, i.e., the quality evaluation result, is obtained. The quality evaluation result is "qualified" or "unqualified".

[0098] The quality assessment method for equipment spare parts of the present invention further trains a feature dimension reduction model and a classifier model based on a trained feature extraction model, thereby constructing a target assessment model for quality assessment of equipment spare parts sample data, reducing the cost of manual operation, and having multiple beneficial effects such as high efficiency, accuracy, reliability, strong generalization ability, high degree of automation, and support for rapid response.

[0099] In some embodiments, the step of training the feature dimensionality reduction model based on the target feature set includes: performing a preset feature dimensionality reduction process on the target feature set based on the feature dimensionality reduction model to generate a first feature dimensionality reduction set. Based on the first feature dimensionality reduction set, determining a second loss value of the feature dimensionality reduction model. Based on the second loss value, updating a second network parameter of the feature dimensionality reduction model, and returning to the step of performing a preset feature dimensionality reduction process on the target feature set until the feature dimensionality reduction model meets the second preset training condition, and determining the trained feature dimensionality reduction model.

[0100] In some embodiments, Figure 3As shown in the figure, the data after feature extraction (i.e., the target feature set) is input into the feature dimensionality reduction model for training. The autoencoder algorithm based on inverse reconstruction is used as the feature dimensionality reduction model. The autoencoder includes an encoder and a decoder. The encoder is responsible for compressing the input data into a low-dimensional representation, and the decoder reconstructs the data from this low-dimensional representation.

[0101] In some embodiments, the training process of the feature dimensionality reduction model is described:

[0102] Step 1: Initialize the second network parameters corresponding to the autoencoder. The second network parameters include the weight parameters and bias parameters of the autoencoder. The initialization method can be expressed as:

[0103] ;

[0104] ;

[0105] In the formula, is the initial weight matrix of the autoencoder; is the initial bias parameter of the autoencoder; represents the number of neurons in the input layer, represents the number of neurons in the output layer; Indicated in A uniformly distributed random number in the interval.

[0106] Step 2: The input target feature set is converted into a low-dimensional space representation through the encoder, and the low-dimensional representation is reconstructed into an output with the same dimension as the original target feature set through the decoder. During the forward propagation process, the encoder's calculated low-dimensional output can be expressed as:

[0107] ;

[0108] The reconstructed output of the decoder can be expressed as:

[0109] ;

[0110] In the formula, It is the low-dimensional feature representation output by the encoder, that is, the first feature dimensionality reduction set; is the Sigmoid activation function; and are the weight parameters of the encoder and decoder respectively; and are the bias parameters of the encoder and decoder respectively; is the input data, i.e. the target feature set; is the output of the reconstruction.

[0111] Step 3: Calculate the inverse reconstruction loss of the autoencoder, that is, calculate the second loss value. The calculation method of the second loss value can be expressed as:

[0112] ;

[0113] In the formula, is the second loss value; is the number of samples input in the current batch; represents the L2 norm; is the regularization parameter; is the KL divergence.

[0114] KL divergence is used to measure the original data distribution and reconstructing data distribution The difference between them is used to ensure that the key features of the data are not lost during the dimensionality reduction process. The calculation method is expressed as:

[0115] ;

[0116] In the formula, and The original data and the reconstructed data are The probability distribution of dimensions, is the feature dimension before feature dimensionality reduction.

[0117] Step 4: Calculate the parameter gradient according to the loss function, and use the gradient descent algorithm to update the weight parameters and bias parameters of the encoder and decoder respectively, that is, update the second network parameters. The updated weight parameters and bias parameters of the encoder and decoder are respectively expressed as:

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] In the formula, is the learning rate; is the updated encoder weight parameter; is the updated encoder bias; is the updated decoder weight parameter; is the updated decoder bias.

[0123] Furthermore, the gradient For example, the corresponding calculation method can be expressed as:

[0124] ;

[0125] In the formula, represents the partial derivative of the loss function with respect to the reconstructed output, is the partial derivative during decoding, is the partial derivative during encoding.

[0126] Step 5: Repeat the above steps until the second preset training condition is met, indicating that the feature dimension reduction model training is completed. The second preset training condition is to reach a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1200 times.

[0127] In some embodiments, a second loss value is calculated based on the reconstructed data obtained based on the first feature dimensionality reduction set and the target feature set, and the weight parameters and bias parameters (i.e., second network parameters) of the encoder and decoder are dynamically updated based on the second loss value.

[0128] The quality assessment method for equipment spare parts of the present invention adopts an inverse reconstruction strategy to process the high-dimensional data after input feature extraction, which can enhance the feature dimensionality reduction adaptability of the feature dimensionality reduction model and minimize the dependency between different features, thereby improving the quality of encoding and the accuracy of decoding.

[0129] In some embodiments, the step of training a classifier model based on a target dimensionality reduction set includes: based on the classifier model, performing preset classification processing on each sample data in the target dimensionality reduction set to generate a prediction data set; based on all prediction data sets, determining a third loss value of the classifier model; updating a third network parameter of the classifier model based on the third loss value, and returning to the step of performing preset classification processing on each sample data in the target dimensionality reduction set until the classifier model meets the third preset training condition, thereby determining the trained classifier model.

[0130] In some embodiments, the reduced-dimensionality data (i.e., the target reduced-dimensionality set) is input into a classifier model to train the classifier model. A probabilistic neural network based on a dynamic weight gradient backtracking strategy is used to construct the classifier model, and in order to enhance the adaptability of the classifier model to data changes, a dynamic adjustment mechanism based on feedback is used in the weight update mechanism of the probabilistic neural network. The dynamic adjustment mechanism based on feedback can dynamically adjust the step size of the weight parameter update according to the classification accuracy of each batch of data, thereby optimizing the learning process.

[0131] In some embodiments, the training process of the classifier model is described:

[0132] Step 1: Initialize the weight parameters and bias parameters of all neural network layers in the classifier model. In one embodiment, the initialization method is expressed as:

[0133] ;

[0134] ;

[0135] In the formula, is the weight parameter of the classifier model, is the bias parameter of the classifier model, is the initialization standard deviation of the weight parameters of the classifier model, is the initialization standard deviation of the bias parameters of the classifier model.

[0136] Step 2: In each training iteration, the data input to the classifier model is processed through a sequence of network layers, and the output of each layer is converted into the input of the next layer through an activation function until the last layer produces a classification result. Specifically, let is the input feature vector of the classifier model. The output of the forward propagation is obtained by accumulating the weighted input of each layer and the activation function, which can be expressed as:

[0137] ;

[0138] In the formula, is the predicted probability of the classifier model; is the feature vector in the target dimension reduction set; is the feature vector after self-supervision adjustment; is the Sigmoid activation function.

[0139] In order to improve the accuracy of the classifier in the quality assessment of equipment spare parts, a feature correction mechanism based on self-supervised learning is adopted. For the feature deviation and noise problems that may be encountered in practical applications, self-supervised learning is used to fine-tune and correct the feature representation, thereby making the probabilistic neural network more robust and improving the generalization ability of the probabilistic neural network. Specifically, the self-supervised learning layer is used to adjust the features, and the adjusted feature vector can be expressed as:

[0140] ;

[0141] In the formula, is the learned adjustment range, are the parameters of the self-supervised learning layer, Self-supervised learning layer processing function.

[0142] Furthermore, the self-supervised learning layer is based on the self-supervised autoencoder structure, and its goal is to reconstruct the input features and learn possible deviations at the same time. The calculation of the self-supervised learning layer can be expressed as:

[0143] ;

[0144] In the formula, and They are the self-supervised encoder and self-supervised decoder functions, both parameterized as .

[0145] Furthermore, the optimization goal of self-supervised learning is to minimize the reconstruction error and adjust the input features to adapt to the classification task. The objective function is:

[0146] ;

[0147] Furthermore, a regularization term is used to ensure that the adjustment is not too large and maintain the authenticity of the features, which is expressed as:

[0148] ;

[0149] In the formula, is the amplitude regularization coefficient; is the feature adjustment amplitude, which is learned dynamically through self-supervised learning.

[0150] Step 3: Use the cross entropy loss function to calculate the error between each predicted value in the output prediction data set and the annotation of the actual sample data (ie, the actual label), that is, the third loss value, to evaluate the performance of the model in the current iteration. The calculation method of the third loss value of the classifier model is expressed as:

[0151] ;

[0152] In the formula, is the number of categories for the classification task, is the loss function of the classifier model, For the The true label of the class, For the The true probability of the class is the predicted probability of the neural network.

[0153] The above The calculation method of the predicted probability of the true probability neural network of the class is expressed as:

[0154] ;

[0155] In the formula, It is The output of the last layer of the neural network corresponding to the class, It is The output of the last layer of the neural network corresponding to the class.

[0156] Step 4: Calculate the gradient of each layer weight using the chain rule according to the loss function. The calculation method is expressed as:

[0157] ;

[0158] ;

[0159] In the formula, is the loss function About weight parameters The gradient of is the loss function About bias parameters gradient.

[0160] Step 5: Update the weight parameters and bias parameters of the classifier model, that is, update the third network parameters. The update method can be expressed as:

[0161] ;

[0162] ;

[0163] In the formula, is the learning rate, is a dynamic factor adjusted according to the classification accuracy, used to control the learning step size. is the weight parameter of the updated classifier model, is the bias parameter of the updated classifier model. Preferably, Set to 0.01.

[0164] Furthermore, the dynamic factor adjusted according to the classification accuracy is adjusted according to the current performance of the model, and the adjustment method can be expressed as:

[0165] ;

[0166] In the formula, is a dynamic factor hyperparameter, is the accuracy of the classifier model of the current batch in predicting the data set. This formula ensures that the learning rate is increased when the accuracy is lower than 50%, and the learning rate is reduced when it is higher than 50%, thereby dynamically adjusting the learning process. Preferably, Set to 0.3.

[0167] Step 6: Repeat the above steps until the third preset training condition is met, indicating that the feature dimension reduction model training is completed. The third preset training condition is to reach a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0168] The quality assessment method for equipment spare parts of the present invention constructs a classifier model based on a probabilistic neural network with a dynamic weight gradient backtracking strategy, and adopts a dynamic adjustment mechanism based on feedback in the probabilistic neural network to dynamically adjust the step size of weight update according to the classification accuracy of each batch of predicted data, thereby optimizing the classifier model learning process.

[0169] In some embodiments, the step of obtaining a target sample data set of equipment spare parts includes: obtaining an original sample data set of equipment spare parts, labeling each sample data in the original sample data set, and the labeling categories include qualified and unqualified; based on the data expansion model, performing data expansion processing on the labeled original sample data set to generate a first expanded data set; obtaining the true probability corresponding to the first expanded data set, updating the fourth network parameter of the data expansion model based on the true probability, and returning to the step of performing data expansion processing on the labeled original sample data set until the data expansion model meets the fourth preset training condition, and determining the trained data expansion model; based on the trained data expansion model, performing data expansion processing on the labeled original sample data set to generate a second expanded data set, and obtaining the target sample data set based on the second expanded data set.

[0170] In some embodiments, the sample data in the original sample data set comprehensively covers the entire process of equipment spare parts from production to inspection. These data are directly collected from precision sensors and quality inspection equipment distributed in each production line in a highly automated manner. The sample data is transmitted to the central database in real time and efficiently using seamless network connections between devices, ensuring the timeliness and accuracy of the data. All sample data can be stored in the central database in a structured JSON format. The sample data set obtained from the central database is used as the original sample data set. The data items of the sample data are carefully designed to fully reflect the key attributes of spare parts, including but not limited to (Material type, such as metal, plastic, etc.), (component size, such as in millimeters), (weight, e.g. grams), (color, following the RGB color space standard), (Hardness, following Mohs hardness standard), (elastic modulus, in Pascals), (thermal conductivity, unit: W / m•K), (resistivity in ohm-meters) and (Melting point, unit: degrees Celsius). It should be emphasized that the above is only an example of the sample data format and attributes of the present invention. In actual applications, the richness of the key attributes of the sample data is far beyond this, and the number of attributes may be expanded to dozens or even hundreds of items to fully cover the various physical and chemical properties of spare parts. The collected original sample data set is labeled, and the labeling method can be manual labeling, and the labeled categories include "qualified" and "unqualified".

[0171] In some embodiments, in the field of equipment spare parts quality assessment, comprehensive collection, accurate labeling and efficient preprocessing of training data are extremely challenging tasks, which are not only time-consuming and laborious, but also often limit the generalization ability and prediction accuracy of each model due to the scarcity of training samples. In order to overcome this bottleneck, the present invention introduces a generative adversarial network (GAN) based on nonlinear data flow as a data expansion model to generate high-quality sample data in an automated manner to achieve effective data expansion. The data expansion model includes a generator and a discriminator. The generator is responsible for generating sample data that is as realistic as possible from the latent space, and the discriminator is responsible for distinguishing the sample data generated by the generator from the real sample data. Furthermore, on the basis of the generative adversarial network, an asymmetric information flow coupling optimization mechanism is adopted to optimize the diversity and authenticity of the generated sample data by enhancing the information exchange between the generator and the discriminator, so that the generated data better reflects the intrinsic characteristic distribution of the real data. The fourth network parameter is the weight parameter and bias parameter of the generator and the discriminator.

[0172] In some implementations, the training process of the data augmentation model is described:

[0173] Step 1: Initialize the network parameters of the generator and discriminator in the data augmentation model. The weight parameters and bias parameters of the generator and discriminator are initialized using a random initialization method;

[0174] Step 2: Sample the input vector from the multidimensional Gaussian distribution (high-dimensional latent space) as the input of the generator. The multidimensional Gaussian distribution (also known as the multivariate normal distribution) is defined by the mean vector and the covariance matrix. Use a random number generator to sample the input vector from the defined multidimensional Gaussian distribution. The input vector can be expressed as:

[0175] ;

[0176] In the formula, represents the input vector of the generator; represents the mean vector of a multidimensional Gaussian distribution; represents the variance of the normal distribution; represents the identity matrix; is a normal distribution. Preferably, Set to 0, Set to 1. The identity matrix is ​​a square matrix with all 1s on the diagonal and all 0s elsewhere.

[0177] Step 3: The generator receives the input vector and generates sample data, maps the input vector to the data space, and forms a fake data sample. Generally, the sample data generated by the generator can be expressed as:

[0178] ;

[0179] In the formula, Represents the generated data sample, that is, the fake data sample; Represents the parameters of the generator, including weight parameters and bias parameters; represents a generator function, Represents a generator; To generate enhanced noise for sample data.

[0180] Furthermore, the enhanced noise of the generated sample data The calculation method can be expressed as:

[0181] ;

[0182] In the formula, To enhance the noise adjustment coefficient, it is used to classify the discriminator according to its classification accuracy. Adjust the noise intensity; is an exponential function that ensures that the noise level gradually decreases as the accuracy increases, so as to carefully control the balance between exploration and exploitation in the generation process.

[0183] Based on the above, the generator can be expressed as follows using a convolutional neural network-based structure to generate fake sample data:

[0184] ;

[0185] In the formula, Represents the LeakyReLU activation function; Indicates the number of convolutional layers; Indicates The weight parameters of the convolutional layers; Indicates The bias of the convolutional layer; represents a convolution operation. Preferably, Set to 3. Lea(⋅) is applied to The accumulated results of linear transformations are used to introduce nonlinearity and help the generator learn complex data distributions. Data augmentation is performed based on the above expression, and the first augmented data set is constructed for each fake data sample generated in each iteration.

[0186] Step 4: The discriminator also receives fake sample data from the generator (i.e., the sample data in the first augmented dataset) and the real sample data (i.e., sample data in the original sample data set) and tries to distinguish which are the real sample data collected and which are the fake sample data generated by the generator. The generator tries to deceive the discriminator into classifying the fake sample data as real sample data. Evaluating on real sample data and fake sample data The way is expressed as:

[0187] ;

[0188] In the formula, is the discriminator function, is the discriminator; Represents the discriminator for the sample The output of represents the probability that the sample is considered to be a real data sample, that is, the true probability; Represents the weight parameter of the discriminator; represents the bias of the discriminator; Represents an input sample, which may be or .

[0189] Step 5: At each iteration, the parameter update strategy of the generator and the discriminator is adjusted through the asymmetric information flow coupling mechanism, that is, the fourth network parameters are updated to ensure that information flows effectively between the two networks and enhance the quality and diversity of the generated samples. The update method of the fourth network parameters is expressed as:

[0190] ;

[0191] ;

[0192] ;

[0193] In the formula, Represents the updated generator parameters; Represents the learning rate of the data augmentation model; Express The gradient of Represents the updated discriminator weight parameters; represents the updated discriminator bias parameter; represents the discriminator learning rate; right The gradient of Express The gradient of represents the momentum factor, which is used to accelerate the convergence of the data augmentation model learning process while maintaining a certain influence of past gradients; represents the updated value of the generator parameter in the previous iteration, Represents the updated value of the discriminator weight parameter in the previous iteration; Represents the updated value of the discriminator bias parameter in the previous iteration; is the feature similarity comparison loss. Preferably, Set to 0.01, Set to 0.3, Set to 0.4.

[0194] Step 6: The contrast loss used helps the generator learn how to generate fake sample data that is closer to the real sample data at the feature level. By calculating the feature similarity between the fake sample data and the real sample data, the fake sample data generated by the generator is similar to the real sample data in terms of features. The calculation method is expressed as:

[0195] ;

[0196] In the formula, represents the contrast loss function; represents the coefficient that controls the balance between similarity and discriminator output; represents the L2 norm; Represents the number of sample data in the first augmented data set at each iteration; and Represent the generated fake sample data and real sample data respectively; represents the true probability of the i-th sample data in the first expanded data set. Preferably, Set to 0.3.

[0197] Step 7: Repeat the above steps until the fourth preset training condition is met, indicating that the model training is completed. The fourth preset training condition is to reach a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0198] In some embodiments, after the data expansion model is trained, the trained data expansion model is used to perform data expansion processing on the sample data. Assuming that the original sample data set includes 800 sample data, and the trained data expansion model expands and generates 200 sample data (i.e., the second expanded data set), the expanded data set includes 1000 sample data, that is, the sample data of the second expanded data set is added to the sample data of the original sample data set, and the target sample data set including 1000 sample data is obtained.

[0199] The quality assessment method for equipment spare parts of the present invention, through automated data collection and labeling, uses a structured data format to store a variety of attribute information, including material type, component size, weight, etc., and further adopts a generative adversarial network based on nonlinear data flow for data expansion, which solves the time-consuming and labor-intensive problems of training data collection, labeling and preprocessing, avoids the problems of insufficient sample quantity and data sparsity, and enhances the generalization ability of the model.

[0200] Based on the same inventive concept, according to another aspect of the present invention, Figure 4 As shown, an embodiment of the present invention further provides a quality assessment system for equipment spare parts, including:

[0201] An acquisition unit 110 is configured to acquire a target sample data set of equipment spare parts;

[0202] The feature extraction unit 120 is configured to construct a feature extraction model based on a preset fully connected neural network, and perform preset feature extraction processing on the target sample data set based on the feature extraction model to obtain a first feature set;

[0203] The training unit 130 is configured to update the first network parameter of the feature extraction model based on the first feature set, and return to the feature extraction unit 120 to perform preset feature extraction processing on the target sample data set again until the feature extraction model meets the first preset training condition;

[0204] The evaluation unit 140 is configured to construct a target evaluation model based on the trained feature extraction model to perform quality evaluation on the sample data of equipment spare parts to be evaluated.

[0205] The quality assessment system for equipment spare parts of the present invention encapsulates a feature extraction model based on a 6-layer fully connected neural network structure, and learns and extracts deep, nonlinear features in the data through the output layer and multiple hidden layers, which helps to capture subtle differences and complex patterns that are difficult to find with traditional methods, thereby improving the accuracy and reliability of subsequent quality assessments. An exploratory gradient descent method is used to update the first network parameters of the feature extraction model until preset training conditions are met, so that the feature extraction model can continuously learn and improve from the data, avoiding the problems of gradient vanishing, gradient explosion or falling into a local optimal solution, and ultimately obtaining a more accurate and generalized feature representation, which helps to improve the robustness of the assessment model. A target assessment model is constructed based on the trained feature extraction model to perform quality assessment of equipment spare parts sample data, thereby reducing the cost of manual operation, and having multiple beneficial effects such as high efficiency, accuracy, reliability, strong generalization ability, high degree of automation, and support for rapid response.

[0206] Based on the same inventive concept, according to another aspect of the present invention, Figure 5 As shown, an embodiment of the present invention further provides a computer device 30, which includes a processor 310 and a memory 320. The memory 320 stores a computer program 321 that can be run on the processor. When the processor 310 executes the program, the steps of the above method are performed.

[0207] Based on the same inventive concept, according to another aspect of the present invention, Figure 6 As shown, an embodiment of the present invention further provides a computer-readable storage medium 40, which stores a computer program 410 for executing the above method when executed by a processor.

[0208] Finally, it should be noted that a person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium of the program can be a disk, an optical disk, a read-only storage memory (ROM) or a random access memory (RAM), etc. The above-mentioned computer program embodiments can achieve the same or similar effects as the corresponding above-mentioned any method embodiments.

[0209] It will also be appreciated by those skilled in the art that various exemplary logic blocks, modules, circuits and algorithm steps described in conjunction with the disclosure herein can be implemented as electronic hardware, computer software or a combination of the two. In order to clearly illustrate this interchangeability of hardware and software, a general description has been given to the functions of various schematic components, blocks, modules, circuits and steps. Whether this function is implemented as software or hardware depends on specific applications and the design constraints imposed on the entire system. Those skilled in the art can implement the function in various ways for each specific application, but this implementation decision should not be interpreted as causing a departure from the disclosed scope of the embodiments of the present invention.

[0210] The above are exemplary embodiments disclosed in the present invention, but it should be noted that various changes and modifications may be made without departing from the scope of the embodiments disclosed in the present invention as defined in the claims. The functions, steps and / or actions of the method claims according to the disclosed embodiments described herein do not need to be performed in any particular order. The serial numbers of the embodiments disclosed in the above embodiments of the present invention are for description only and do not represent the advantages and disadvantages of the embodiments. In addition, although the elements disclosed in the embodiments of the present invention may be described or required in individual form, they may also be understood as multiple unless explicitly limited to the singular.

[0211] It should be understood that, as used herein, the singular forms "a", "an" are intended to include the plural forms as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and / or" refers to any and all possible combinations including one or more of the associated listed items.

[0212] A person skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples; under the concept of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other changes in different aspects of the above embodiments of the present invention, which are not provided in detail for the sake of simplicity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the protection scope of the embodiments of the present invention.

Claims

1. A quality assessment method for equipment spare parts, characterized in that: include: Obtain target sample data set of equipment spare parts; Building a feature extraction model based on a preset fully connected neural network, and performing preset feature extraction processing on the target sample data set based on the feature extraction model to obtain a first feature set; Based on the first feature set, a first network parameter of the feature extraction model is updated, and the step of performing a preset feature extraction process on the target sample data set is returned until the feature extraction model meets a first preset training condition; Based on the trained feature extraction model, a target evaluation model is constructed to perform quality evaluation on the sample data of equipment spare parts to be evaluated; The step of constructing a feature extraction model based on a preset fully connected neural network includes: On the basis of the preset fully connected neural network encapsulated by an input layer, an output layer and four hidden layers, constructing the feature extraction model based on an exploratory gradient descent strategy; The step of performing preset feature extraction processing on the target sample data set based on the feature extraction model to obtain a first feature set includes: In response to the input layer in the feature extraction model receiving the target sample data set, the target sample data set is used as target input data, and the next neural network layer of the input layer is used as the target network layer; Performing a preset linear change process on the target input data based on the target network layer, and performing a preset feature extraction process on the target input data after the linear change based on a ReLU activation function, so as to obtain a feature result output by the target network layer; Using the feature result as new target input data, using the next neural network layer of the target network layer as the new target network layer, and returning to the step of performing the preset linear change processing on the target input data based on the target network layer until there is no next neural network layer in the target network layer, and using the feature result as the first feature set; Wherein, the next neural network layer includes a hidden layer or an output layer; The step of updating the first network parameter of the feature extraction model based on the first feature set comprises: Determining a first loss value of the output layer according to the first feature set; Based on the first loss value, determine the first weight gradient value and the first bias gradient value corresponding to each of the hidden layers respectively; Based on each of the first weight gradient values ​​and each of the first bias gradient values, predicting a second weight gradient value and a second bias gradient value respectively corresponding to each of the hidden layers; Based on each of the first weight gradient values, each of the first bias gradient values, each of the second weight gradient values, and each of the second bias gradient values, the weight parameters and bias parameters corresponding to each of the hidden layers are updated to update the first network parameters of the feature extraction model.

2. The quality assessment method for equipment spare parts according to claim 1 is characterized in that: The step of constructing a target evaluation model based on the trained feature extraction model includes: Obtaining a target feature set corresponding to the target sample data set output by the trained feature extraction model; Based on the target feature set, a feature dimensionality reduction model is trained, and a target dimensionality reduction set corresponding to the target feature set output by the trained feature dimensionality reduction model is obtained; Based on the target dimension reduction set, training a classifier model; A target evaluation model is constructed based on the trained feature extraction model, the trained feature dimension reduction model and the trained classifier model.

3. The quality assessment method for equipment spare parts according to claim 2 is characterized in that: The step of training a feature dimensionality reduction model based on the target feature set includes: Based on the feature dimensionality reduction model, performing a preset feature dimensionality reduction process on the target feature set to generate a first feature dimensionality reduction set; Determining a second loss value of the feature dimensionality reduction model based on the first feature dimensionality reduction set; Based on the second loss value, the second network parameter of the feature dimensionality reduction model is updated, and the step of performing a preset feature dimensionality reduction process on the target feature set is returned until the feature dimensionality reduction model meets the second preset training condition, thereby determining the trained feature dimensionality reduction model.

4. The quality assessment method for equipment spare parts according to claim 2 is characterized in that: The step of training a classifier model based on the target dimension reduction set includes: Based on the classifier model, each sample data in the target dimension reduction set is subjected to a preset classification process to generate a prediction data set; Determining a third loss value of the classifier model based on all predicted data sets; Based on the third loss value, the third network parameter of the classifier model is updated, and the step of performing preset classification processing on each sample data in the target dimensionality reduction set is returned until the classifier model meets the third preset training condition, thereby determining the trained classifier model.

5. The quality assessment method for equipment spare parts according to claim 1 is characterized in that: The step of obtaining a target sample data set of equipment spare parts includes: Acquire an original sample data set of the equipment spare parts, and mark each sample data in the original sample data set, wherein the marked categories include qualified and unqualified; Based on the data expansion model, performing data expansion processing on the annotated original sample data set to generate a first expanded data set; Obtaining a true probability corresponding to the first expanded data set, updating a fourth network parameter of the data expansion model based on the true probability, and returning to the step of performing data expansion processing on the labeled original sample data set until the data expansion model meets a fourth preset training condition, and determining the trained data expansion model; Based on the trained data expansion model, data expansion processing is performed on the labeled original sample data set to generate a second expanded data set, and based on the second expanded data set, the target sample data set is obtained.

6. A quality assessment system for equipment spare parts, characterized in that: include: An acquisition unit configured to acquire a target sample data set of equipment spare parts; A feature extraction unit is configured to construct a feature extraction model based on a preset fully connected neural network, and perform preset feature extraction processing on the target sample data set based on the feature extraction model to obtain a first feature set; A training unit configured to update a first network parameter of the feature extraction model based on the first feature set, and return to the feature extraction unit to perform a preset feature extraction process on the target sample data set again until the feature extraction model meets a first preset training condition; An evaluation unit configured to construct a target evaluation model based on the trained feature extraction model to perform quality evaluation on the equipment spare parts sample data to be evaluated; The feature extraction unit is further configured to: construct the feature extraction model based on an exploratory gradient descent strategy on the basis of the preset fully connected neural network encapsulated by an input layer, an output layer and four hidden layers; The feature extraction unit is further configured to: in response to the input layer in the feature extraction model receiving the target sample data set, use the target sample data set as target input data, and use the next neural network layer of the input layer as the target network layer; Performing a preset linear change process on the target input data based on the target network layer, and performing a preset feature extraction process on the target input data after the linear change based on a ReLU activation function, so as to obtain a feature result output by the target network layer; The feature result is used as new target input data, the next neural network layer of the target network layer is used as the new target network layer, and the step of performing the preset linear change processing on the target input data based on the target network layer is returned until the target network layer does not have a next neural network layer, and the feature result is used as the first feature set; wherein the next neural network layer includes a hidden layer or an output layer; The training unit is also configured to: determine the first loss value of the output layer according to the first feature set; determine the first weight gradient value and the first bias gradient value corresponding to each of the hidden layers based on the first loss value; predict the second weight gradient value and the second bias gradient value corresponding to each of the hidden layers based on each of the first weight gradient values ​​and each of the first bias gradient values; update the weight parameters and bias parameters corresponding to each of the hidden layers based on each of the first weight gradient values, each of the first bias gradient values, each of the second weight gradient values, and each of the second bias gradient values, so as to update the first network parameters of the feature extraction model.

7. A computer device comprising: at least one processor; as well as A memory storing a computer program executable on the processor, wherein the processor executes the steps of the method according to any one of claims 1 to 5 when executing the program.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are performed.

Images