Artificial Intelligence-based Enterprise Digital Management Maturity Assessment Method

Abstract

The present invention relates to the field of artificial intelligence technology, and specifically relates to a method for evaluating the maturity of enterprise digital management based on artificial intelligence. The method uses a generative adversarial network algorithm based on a variant of momentum optimization to generate samples and achieve data augmentation. The augmented enterprise management data is input into a feature extraction model for training the feature extraction model. The enterprise management data after feature extraction is input into a feature dimensionality reduction model for training the feature dimensionality reduction model. The enterprise management data after dimensionality reduction is input into a classifier model for training the classifier. The trained model is used to process new sample data to evaluate the digital management maturity level it is in.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an evaluation method for the maturity of enterprise digital management based on artificial intelligence. Background Art

[0002] With the rapid development of information technology, enterprises are gradually implementing digital transformation to improve operational efficiency and market competitiveness. Digital management not only involves the application of technology, but also includes a comprehensive transformation of corporate culture, strategic decision-making, and internal processes. Therefore, accurately evaluating the maturity of enterprises in the process of digital transformation has become particularly important, which helps enterprises understand the current situation, formulate reasonable development strategies, and optimize management practices.

[0003] The existing technologies mainly rely on traditional data processing methods and statistical analysis to evaluate the maturity of enterprise digital management. These methods often rely on the subjective judgment of experts and simple quantitative analysis, lacking in-depth mining of large-scale complex data and dynamic adaptation capabilities. In addition, traditional methods have many limitations in data processing and model training, such as insufficient model generalization ability caused by limited data volume, and serious information loss in feature extraction and dimensionality reduction processing.

[0004] These deficiencies affect the accuracy and efficiency of the evaluation, and limit the ability of enterprises to make timely and effective decisions in a rapidly changing market environment. Summary of the Invention

[0005] The purpose of the present invention is to provide an evaluation method for the maturity of enterprise digital management based on artificial intelligence, which solves the problem that the deficiencies of the existing enterprise digital management affect the accuracy and efficiency of the evaluation, and limit the ability of enterprises to make timely and effective decisions in a rapidly changing market environment.

[0006] To achieve the above purpose, the present invention provides an evaluation method for the maturity of enterprise digital management based on artificial intelligence, including the following steps:

[0007] Collect enterprise management data and perform annotation;

[0008] Adopt a generative adversarial network algorithm based on a variant of momentum optimization for sample generation to obtain an expansion of enterprise management data;

[0009] Input the expanded enterprise management data into a feature extraction model for training the feature extraction model;

[0010] Input the enterprise management data after feature extraction into a feature dimensionality reduction model for training the feature dimensionality reduction model;

[0011] Input the enterprise management data after dimensionality reduction into a classifier model for training the classifier;

[0012] Use the trained model to process new sample data to evaluate its digital management maturity level.

[0013] Among them, the generative adversarial network algorithm based on a variant of momentum optimization is used for sample generation to obtain enterprise management data augmentation. The steps also include:

[0014] The generative adversarial network algorithm based on a variant of momentum optimization includes a generator and a discriminator. Among them, the goal of the generator is to generate as realistic enterprise management data as possible to deceive the discriminator; the task of the discriminator is to distinguish between real enterprise management data and the fake enterprise management data generated by the generator.

[0015] Among them, the training process of the generative adversarial network algorithm based on a variant of momentum optimization includes:

[0016] Initialize the parameters of the generative adversarial network model based on a variant of momentum optimization;

[0017] Use the Adam optimizer with a dynamic learning rate to optimize the parameters of the generator and the discriminator;

[0018] In each training cycle, the generator generates a batch of fake enterprise management data, and the discriminator evaluates the difference between these enterprise management data and the real enterprise management data. The feedback of the discriminator is used to adjust the parameters of the generator to make the enterprise management data it generates more realistic;

[0019] Use the adaptive feature enhancement technology to optimize the quality of the generated enterprise management data. By dynamically adjusting the key features in the generated enterprise management data, according to the feedback of the discriminator, to enhance the adaptability and accuracy of the model to complex enterprise management scenarios;

[0020] According to the feedback of the discriminator, continuously adjust the strategy for generating enterprise management data to better adapt to the changes in the distribution of enterprise management data. The adjustment of the generation strategy is achieved by means of an adaptive mechanism;

[0021] After a fixed period, evaluate the generation effect of the model and update the parameters of the generator and the discriminator accordingly;

[0022] Repeat the above steps iteratively until the preset stop iteration condition is met, which means the model training is completed.

[0023] Among them, input the augmented enterprise management data into the feature extraction model for training the feature extraction model. The steps also include:

[0024] Initialize the neural network model;

[0025] In the growth period, quickly adapt to the characteristics of enterprise management data by accelerating the fine-tuning of weights;

[0026] During the synthesis period, weight synthesis is performed;

[0027] During the evaluation period, the current network performance is evaluated and a decision is made on whether to adjust the cycle optimization strategy. If so, return to the growth period; if not, proceed to the pre - division stage;

[0028] During the pre - division stage, simulate the cell's preparatory division stage and make fine adjustments to the weights;

[0029] During the division period, implement structural adjustment;

[0030] During the dormancy period, adjust the learning rate and perform learning rate decay;

[0031] Repeat the above steps iteratively until the preset stop - iteration condition is met, indicating that the model training is complete.

[0032] Among them, the enterprise management data after feature extraction is input into the feature dimensionality reduction model for training the feature dimensionality reduction model. The steps also include:

[0033] Initialize all the weights of the auto - encoder;

[0034] During the encoder training stage, input the enterprise management data after feature extraction and convert it into a low - dimensional feature representation. Each input enterprise management data is gradually compressed through the encoder hierarchy. Each layer of the encoder calculates a new output through the output of the previous layer and the weights of the current layer;

[0035] After the features are encoded, adjust the network weights according to the inherent ambiguity of the enterprise management data. Evaluate the importance and contribution degree of each feature through fuzzy logic and dynamically adjust the weights according to the evaluation results;

[0036] The decoder receives the low - dimensional features output by the encoder and attempts to reconstruct the original input data. Each low - dimensional feature is expanded layer by layer until data of the same dimension as the original input is reconstructed;

[0037] By comparing the differences between the reconstructed data and the original input data, calculate the error and update the weights in the network through the back - propagation algorithm to reduce the reconstruction error;

[0038] Adopt an adaptive feature attenuation strategy to optimize the dimensionality reduction quality and reconstruction accuracy of enterprise management data by dynamically adjusting the retention or attenuation degree of features in the decoding stage during the training process of the feature dimensionality reduction model of the auto - encoder;

[0039] Repeat the above steps iteratively until the preset stop - iteration condition is met, indicating that the model training is complete.

[0040] Among them, the dimension-reduced enterprise management data is input into a classifier model for training the classifier, and the steps further include:

[0041] In the initialization stage, the weights and biases from the input layer to the hidden layer are randomly initialized;

[0042] Calculate the representation of each sample in the hidden layer;

[0043] Use the least squares method to adjust the output layer weights and minimize the difference between the actual output and the expected output;

[0044] Adopt fractional gradient adjustment to optimize the weights by calculating the fractional-order derivative of the gradient;

[0045] Repeat the above steps iteratively until the preset stop iteration condition is met, which indicates that the model training is completed.

[0046] An enterprise digital management maturity evaluation method based on artificial intelligence of the present invention adopts a generative adversarial network algorithm based on momentum optimization variant for data augmentation. Through the collaborative work of the generator and the discriminator, realistic enterprise management data can be effectively generated, which can overcome the problem of insufficient training caused by limited data samples, improve the generalization ability of the model, enhance the diversity and authenticity of the generated data, overcome the problem of model overfitting caused by insufficient data volume, and improve the prediction accuracy of the model.

[0047] Adopt a circular dynamic optimization algorithm to optimize the parameters of the neural network, simulate different stages in the cell cycle, and enhance the adaptability and accuracy of the model to complex data environments by adjusting the optimization strategy stage by stage, effectively reducing the risk of overfitting, and being able to effectively jump out of the local optimal solution, thereby improving the accuracy and efficiency of feature extraction.

[0048] Adopt an autoencoder based on lightweight fuzzy weights to perform dimension reduction processing on enterprise management data. By dynamically adjusting the fuzzy logic weights in the weight update process, the accuracy of the dimension-reduced data and the generalization ability of the network are improved. By adjusting the fuzzy logic weights, the autoencoder not only reduces the reconstruction error, but also can more accurately reflect the structural characteristics of the original data, enhancing the stability and reliability of data processing.

[0049] Adopt an extreme learning machine algorithm based on fractional gradient adjustment to improve the classifier training process. Through the meticulous control of the fractional-order derivative, the processing ability of nonlinear problems and the generalization of the classifier are optimized. The fractional gradient adjustment enables the extreme learning machine to more finely adjust the weights during the learning process, reduces the risk of model overfitting, and improves the ability to handle complex classification problems. Description of the Drawings

[0050] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art.

[0051] Figure 1 is a flowchart of the method for evaluating the maturity of enterprise digital management based on artificial intelligence in the first embodiment of the present invention.

[0052] Figure 2 is a training flowchart of the neural network algorithm based on circular dynamic optimization in the first embodiment of the present invention.

[0053] Figure 3 is a flowchart of model training and application in the first embodiment of the present invention.

[0054] Figure 4 is a step diagram of the method for evaluating the maturity of enterprise digital management based on artificial intelligence in the first embodiment of the present invention. Detailed implementation manners

[0055] The following will describe in detail the embodiments of the present invention. The examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present invention, and should not be construed as a limitation to the present invention.

[0056] The first embodiment of the present application is as follows:

[0057] Please refer to Figures 1 to 4 , where Figure 1 is a flowchart of the method for evaluating the maturity of enterprise digital management based on artificial intelligence in the first embodiment of the present invention. Figure 2 is a training flowchart of the neural network algorithm based on circular dynamic optimization in the first embodiment of the present invention. Figure 3 is a flowchart of model training and application in the first embodiment of the present invention. Figure 4 is a step diagram of the method for evaluating the maturity of enterprise digital management based on artificial intelligence in the first embodiment of the present invention.

[0058] The present invention provides a method for evaluating the maturity of enterprise digital management based on artificial intelligence, including the following steps:

[0059] S100: Collect enterprise management data and perform annotation;

[0060] Specifically, the data of the present invention is sourced from the enterprise's internal management systems, such as ERP systems, CRM systems, and human resource management platforms. The enterprise's internal management systems provide a large amount of raw data generated during the enterprise's operation, and the collected data is stored in JSON format. In one embodiment, the attributes of the data include: ax1 is the total assets of the enterprise (numerical); ax2 is the annual operating income (numerical); ax3 is the total number of employees (numerical); ax4 is the proportion of IT department employees (numerical); ax5 is the annual IT investment (numerical); ax6 is the service life of the ERP system (numerical); ax7 is the service life of the CRM system (numerical); ax8 is the cloud service adoption rate (percentage); ax9 is the data security investment ratio (percentage); ax10 is the customer satisfaction score (numerical).

[0061] It should be noted that this embodiment is only to illustrate a data format and type of the present invention. In actual applications, the attributes of enterprise management data usually exceed 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0062] Furthermore, the collected enterprise management data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeling categories include: "high maturity", "medium maturity", and "low maturity", a total of 3 categories.

[0063] S200: Use the generative adversarial network algorithm based on the momentum optimization variant to generate samples and obtain the expansion of enterprise management data;

[0064] Specifically, it can be understood that in the task of the present invention, the collection, acquisition, labeling, and preprocessing of training data are time-consuming and laborious, and insufficient training samples are likely to lead to poor generalization ability of the model and affect the accuracy of the model. The present invention uses the generative adversarial network algorithm based on the momentum optimization variant to generate samples, thereby realizing the expansion of enterprise management data. The generative adversarial network algorithm based on the momentum optimization variant includes a generator (G) and a discriminator (D). The goal of the generator is to generate as realistic enterprise management data as possible to deceive the discriminator; the task of the discriminator is to distinguish between real enterprise management data and the fake enterprise management data generated by the generator. The present invention adopts a constrained optimization strategy in the traditional generative adversarial network to strengthen the quality of the generated enterprise management data, allowing the model to better adapt to various types of enterprise management data distributions, so that the generated enterprise management data has significant improvements in both diversity and authenticity.

[0065] The training process of the generative adversarial network algorithm based on the momentum optimization variant is as follows:

[0066] 1. Initialize the parameters of the generative adversarial network model based on the momentum optimization variant. In one embodiment, the initialization method uses the Gaussian distribution strategy for initialization, which is expressed as:

[0067]

[0068]

[0069] In the formula, represents the initial weight of the generator; represents the initial weight of the discriminator; and are the number of nodes in the input layer and the output layer respectively; is the Gaussian distribution function.

[0070] 2. Use the Adam optimizer with a dynamic learning rate to optimize the parameters of the generator and the discriminator. The optimization method is:

[0071]

[0072]

[0073] In the formula, and represent the parameters of the generator and the discriminator respectively; is the learning rate of the th iteration; and are the first-order and second-order moment estimates; is a small positive number added to prevent division by zero. Preferably, is set to 0.0001; is set using the simulated annealing strategy, that is, as the number of iterations increases, the learning rate decreases at a ratio of 0.95.

[0074] Furthermore, the calculation methods of the first-order moment estimate and the second-order moment estimate are expressed as:

[0075]

[0076]

[0077] In the formula, and are the first-order and second-order moment estimates of the previous iteration respectively; is the gradient at the th iteration; and are the decay rate parameters. Preferably, and Both are set to 0.85.

[0078] 3. In each training cycle, the generator generates a batch of fake enterprise management data, and the discriminator evaluates the differences between this enterprise management data and the real enterprise management data. The feedback of the discriminator is used to adjust the parameters of the generator to make the generated enterprise management data more realistic. Specifically, the ways of generating enterprise management data and evaluation are expressed as:

[0079]

[0080]

[0081]

[0082] In the formula, is the output of the discriminator, is the generated enterprise management data after feature enhancement, representing the probability of being judged as real enterprise management data; is the enterprise management data generated by the generator according to the input and the parameter ; is the function of the discriminator; and are the loss functions of the generator and the discriminator respectively; is the label, 1 for real enterprise management data and 0 for generated enterprise management data.

[0083] 4. The quality of the generated enterprise management data is optimized by using the adaptive feature enhancement technology. By dynamically adjusting the key features in the generated enterprise management data, according to the feedback of the discriminator, the adaptability and accuracy of the model to complex enterprise management scenarios are enhanced. Specifically, the feature difference evaluation function is defined as a function for quantifying the key feature differences between the generated enterprise management data and the real enterprise management data, and the calculation method is:

[0084]

[0085] In the formula, and represent the feature vectors of the generated enterprise management data and the real enterprise management data respectively; is the number of features; is the importance weight of the th feature; : the kth feature of the feature vector of the generated enterprise management data, : the kth feature of the feature vector of the real enterprise management data.

[0086] Furthermore, based on the feature difference evaluation, the adaptive feature enhancement strategy function The calculation method is expressed as:

[0087]

[0088] In the formula, is the learning rate; Indicates the generation of enterprise management data The gradient of the feature difference evaluation function. Preferably, Set to 0.01.

[0089] In one embodiment, the feature importance weights The update strategy is expressed as:

[0090]

[0091] In the formula, is a smoothing parameter. Preferably, Set to 0.5.

[0092] 5. According to the feedback from the discriminator, the strategy for generating enterprise management data is continuously adjusted to better adapt to the changes in the distribution of enterprise management data. The adjustment of the generation strategy is implemented by an adaptive mechanism, which is expressed as:

[0093]

[0094]

[0095] In the formula, is the dynamic adjustment factor of the generator learning rate; is the learning step size; is the adjustment amount of the parameter.

[0096] Furthermore, the learning rate dynamic adjustment factor The calculation method is expressed as:

[0097]

[0098] In the formula, is the decay rate of the adjustment factor; It is in Iterations for parameters The amount of adjustment; represents the second norm of the adjustment amount. Preferably, Set to 0.98.

[0099] 6. After a fixed period, evaluate the generation effect of the model and update the parameters of the generator and discriminator accordingly. The method of updating the parameters is expressed as:

[0100]

[0101]

[0102] In the formula, is the first momentum factor; is the second momentum factor; is the updated generator parameter; is the generator parameter before update; is the updated discriminator parameter; is the discriminator parameter before update. Preferably, the fixed period is set to every 10 iterations.

[0103] In one embodiment, the first momentum factor and the second momentum factor are calculated as follows:

[0104]

[0105]

[0106] In the formula, and are the initial momentum factors; and are the rates at which the momentum factors decay with the number of training iterations; is the current step number of the training iteration. Preferably, is set to 0.9, is set to 0.8.

[0107] 7. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0108] S300: Input the augmented enterprise management data into the feature extraction model for training the feature extraction model;

[0109] Specifically, input the augmented enterprise management data into the feature extraction model for training the feature extraction model. The traditional gradient descent method relies on global gradient information for weight update. Inspired by the periodic metabolic process of biological cells, different mechanisms are used at different stages to optimize the responses and adaptability of organisms.

[0110] The present invention optimizes the parameters of a neural network using a circular dynamic optimization algorithm. The circular dynamic optimization algorithm adjusts the optimization strategy at different training stages, simulating each stage in the cell cycle, and each stage uses a different strategy to adjust the weights and biases of the neural network, enabling the algorithm to more flexibly adapt to the changes in enterprise management data, reducing the risk of overfitting, and effectively jumping out of local optimal solutions.

[0111] The training process of the neural network algorithm based on circular dynamic optimization is as follows:

[0112] 1. Initialize the neural network model. The present invention uses a 3-layer fully connected neural network for feature extraction. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, and the second hidden layer includes 50 neurons. The activation function is the ReL_RU activation function.

[0113] Furthermore, initialize the weights and biases , and set the initial learning rate . In one embodiment, the setting method of the initial weights and biases is expressed as:

[0114]

[0115]

[0116] In the formula, means subject to a specific distribution, is a normal distribution with a mean of 0 and a variance of , is the variance of the initial distribution. Preferably, is set to 0.01.

[0117] 2. In the growth phase, the algorithm quickly adapts to the characteristics of enterprise management data by accelerating the fine-tuning of the weights.

[0118] The present invention uses the accelerated gradient descent algorithm to update the parameters of the neural network. The update method is expressed as:

[0119]

[0120]

[0121] In the formula, is the acceleration coefficient, is the loss function of the neural network training, is the gradient of the loss function with respect to the weight parameter, is the gradient of the loss function with respect to the bias parameter, is the The weight parameter of the th iteration, is the bias parameter of the th iteration, is the weight parameter of the th iteration, is the bias parameter of the is the symbol of the partial derivative, is the learning rate in the growth stage. Preferably, the loss function adopts cross-entropy loss, is set to 0.01.

[0122] Furthermore, the acceleration coefficient is set according to the descent speed and curvature of the loss function to avoid over-adjustment. The calculation method is:

[0123]

[0124] In the formula, is the preset maximum acceleration coefficient, is the function of taking the minimum value.

[0125] 3. In the synthesis stage, the algorithm performs weight synthesis at this stage, that is, synthesizes multiple weight parameters into new parameters through specific operations to improve the generalization ability of the network. In one embodiment, the synthesis function is calculated as:

[0126]

[0127] In the formula, is the importance coefficient of the weight and is the number of weight parameters.

[0128] Furthermore, the importance coefficient is dynamically calculated based on the gradient magnitude in the previous stage . For the th weight parameter, the calculation method of its importance coefficient is expressed as:

[0129]

[0130] In the formula, is the decay speed parameter.

[0131] Furthermore, the calculation method of the decay speed parameter is:

[0132]

[0133] 4. During the evaluation period, the algorithm evaluates the current network performance and decides whether to adjust the cycle optimization strategy. In one embodiment, a performance evaluation function is set up , and it outputs whether to enter the next cycle in the form of a decision function. The decision method is expressed as:

[0134]

[0135] In the formula, is the performance threshold. Preferably, is set to 0.5. If is calculated as 1, it enters the next cycle, that is, enters the growth period; otherwise, it enters the prophase of division.

[0136] 5. In the prophase of division, the preparation stage of cell division is simulated, and the weights are slightly adjusted. In one embodiment, the perturbation method and are used as the adjustment increments of the weights and biases, and the calculation method is expressed as:

[0137]

[0138]

[0139] In the formula, is the adjustment increment of the weight parameter, is the adjustment increment of the bias parameter, and are the amplitude and frequency parameters of fine-tuning. Preferably, is set to 5, is set to 3.14.

[0140] Furthermore, and are set in the following way:

[0141]

[0142]

[0143] In the formula, is the scaling factor of the perturbation amplitude. Preferably, is set to 0.95.

[0144] 6. During the division period, the algorithm implements a large-scale structural adjustment. In one embodiment, the adjustment method is to use weight pruning to constrain the weights. The calculation method of the weight pruning function is expressed as:

[0145]

[0146] In the formula, is the gradient threshold. Preferably, it is set to 0.01, that is, only the parameters with weights greater than 0.01 are retained.

[0147] 7. During the dormancy period, the network gradually reaches a stable state based on the existing weights. In this stage, the learning rate is adjusted so that the learning rate decreases. The adjustment method is expressed as:

[0148]

[0149] In the formula, is the decay coefficient, is the number of completed epochs. Preferably, it is set to 0.95.

[0150] 8. Repeat the above steps iteratively until the preset stop iteration condition is met, which indicates that the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0151] S400: Input the enterprise management data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model;

[0152] Specifically, input the enterprise management data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model. The present invention uses an autoencoder based on lightweight fuzzy weights as the feature dimensionality reduction algorithm. The autoencoder compresses the input enterprise management data into a low-dimensional space through learning and then reconstructs it back to the original space, thereby realizing the dimensionality reduction of the enterprise management data. Moreover, the autoencoder includes an encoder part and a decoder part. The encoder is responsible for compressing the input enterprise management data into low-dimensional features, and the decoder attempts to reconstruct the original input from these low-dimensional features. The present invention improves the autoencoder and uses fuzzy logic weights to adjust the weight update process in the network, dynamically adjusting based on the characteristics of the input enterprise management data, which can better handle the uncertainty and ambiguity in the enterprise management data, thereby improving the accuracy of the dimensionality reduction process and the generalization ability of the network.

[0153] The training process of the autoencoder algorithm based on lightweight fuzzy weights is as follows:

[0154] 1. Initialize all the weights of the autoencoder. In one embodiment, a small random number initialization method is used to break symmetry and avoid the problem of gradient disappearance. The initialization method is expressed as:

[0155]

[0156] In the formula, represents the th weight in the autoencoder; is the standard deviation. Preferably, is set to 0.01.

[0157] 2. During the encoder training stage, the input enterprise management data after feature extraction is transformed into a low-dimensional feature representation. Each input enterprise management data is gradually compressed through the encoder hierarchy. Each layer of the encoder calculates a new output based on the output of the previous layer and the weights of the current layer. The output of each layer of the encoder is calculated as follows:

[0158]

[0159] In the formula, is the output of the th layer; is the input vector or the output of the previous layer; is the weight connecting the th input and the th output; is the bias of the th layer; is the ReLU activation function.

[0160] 3. After the features are encoded, the network weights are adjusted according to the inherent ambiguity of the enterprise management data. The importance and contribution degree of each feature are evaluated through fuzzy logic, and the weights are dynamically adjusted according to these evaluation results. The adjustment method is expressed as:

[0161]

[0162]

[0163] In the formula, is the adjusted weight parameter; is the adjustment amount of the weight; is the weight adjustment factor based on fuzzy logic; is the learning rate. Preferably, is set to 0.01.

[0164] Furthermore, the weight adjustment factor is dynamically calculated according to the relationship between the input and the output . The calculation method is expressed as:

[0165]

[0166] In the formula, is the data input to the i-th layer of the autoencoder; is the output of the i-th layer of the autoencoder; is the fuzziness factor. Preferably, is set to 0.1.

[0167] 4. The decoder accepts the low-dimensional features output by the encoder and then attempts to reconstruct the original input data. Each low-dimensional feature is expanded layer by layer until data with the same dimension as the original input is reconstructed. The output of the decoder is calculated as:

[0168]

[0169] where, is the reconstructed output; are the weights in the decoder; is the bias in the decoder.

[0170] 5. By comparing the differences between the reconstructed data and the original input data, the error is calculated, and the weights in the network are updated through the backpropagation algorithm to reduce the reconstruction error. The update methods of the error and the gradient of the weights are expressed as:

[0171]

[0172]

[0173] where, is the reconstruction error; is the original input data; is the output of the decoder; is the derivative of the ReLU activation function.

[0174] Furthermore, the gradient is calculated as:

[0175]

[0176]

[0177] where, is the weighted sum of the inputs of the -th neuron in the decoder; is the output of the encoder, i.e., the input of the decoder; is the derivative of the activation function at .

[0178] 6. An adaptive feature attenuation strategy is adopted to optimize the dimensionality reduction quality and reconstruction accuracy of enterprise management data by dynamically adjusting the retention or attenuation degree of features in the decoding stage during the training process of the feature dimensionality reduction model of the autoencoder. Specifically, the feature attenuation function For adjusting the decoder output , which is expressed as:

[0179]

[0180] In the formula, is the original output of the decoder; is the attenuation rate parameter; is the attenuation correlation scoring function.

[0181] In one embodiment, the calculation method of the attenuation correlation scoring function is expressed as:

[0182]

[0183] In the formula, is the reconstruction error with respect to the gradient of the decoder output ; is the feature vector of the decoder output.

[0184] In this embodiment, the dynamic adjustment of the attenuation rate is adjusted as follows:

[0185]

[0186] In the formula, is the smoothing factor; is the variance of; is the mean of. Preferably, is set to 0.5.

[0187] 7. Repeat the above steps iteratively until the preset stopping iteration condition is satisfied, which indicates that the model training is completed. In one embodiment, the preset stopping iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times. After the model training is completed, the obtained is the enterprise management data feature after feature dimensionality reduction.

[0188] S500: Input the dimensionality-reduced enterprise management data into the classifier model for classifier training;

[0189] Specifically, input the dimensionality-reduced enterprise management data into the classifier model for classifier training. The present invention uses the extreme learning machine algorithm based on fractional gradient adjustment as the classification algorithm and improves the extreme learning machine. Specifically, the fractional gradient adjustment mechanism is used to more finely control the update of the weights during the learning process of the classifier based on the fractional calculus theory, improving the model's ability to handle nonlinear problems and generalization.

[0190] The training process of the extreme learning machine algorithm based on fractional gradient adjustment is as follows:

[0191] 1. In the initialization stage, the weights from the input layer to the hidden layer and the biases

[0192]

[0193]

[0194] are randomly initialized. In one embodiment, the initialization method is expressed as: where is the standard deviation used for the weights and biases during the initialization process. Preferably,

[0195] 2. Calculate the representation of each sample in the hidden layer. The output of the hidden layer is obtained by linearly combining the input data with the randomly generated weights and biases and then transforming through the non-linear activation function

[0196]

[0197] where is the input data; are the weights of the hidden layer; are the biases of the hidden layer; is the Sigmoid activation function.

[0198] 3. Use the least squares method to adjust the weights of the output layer to minimize the difference between the actual output and the desired output

[0199]

[0200] where is the transpose of the hidden layer output; is the regularization parameter; is the target output; is the identity matrix, with the same dimension as . Preferably, is set to the all-ones matrix.

[0201] 4. Adopt fractional gradient adjustment to optimize the weights by calculating the fractional-order derivative of the gradient, so as to reduce the risk of overfitting while maintaining the learning speed. The update method of the output layer weights is expressed as:

[0202]

[0203]

[0204] In the formula, is the updated output layer weight; is the update increment of the output layer weight; is the error function; represents the fractional derivative of; is the learning rate; is the fractional order parameter.

[0205] Furthermore, the learning rate adopts a dynamic learning rate adjustment mechanism, dynamically adjusts the learning rate by analyzing the error change rate during the training process in real time, so as to optimize the convergence speed and classification accuracy of the model. Specifically, the learning rate is adjusted according to the current gradient of the error function and the change rate of the previous gradient , and the adjustment method is expressed as:

[0206]

[0207] In the formula, is the adjusted and updated learning rate; is the learning rate adjustment factor, usually set to 0.1; function is the sign function; and are the error gradients of the current and previous iterations respectively.

[0208] Furthermore, the calculation method of the error function is expressed as:

[0209]

[0210] In one embodiment, the fractional derivative characterizes the application of fractional calculus to the error function , provides a more detailed weight adjustment than the traditional gradient, and the calculation method is expressed as:

[0211]

[0212] In the formula, is the gamma function, which extends the traditional factorial to the real and complex number fields; is the fractional order parameter, which defines the non-integer number of derivative operations; is the error function with respect to the output The order derivative. Preferably, the fractional-order parameter is set to 0.5.

[0213] 5. Repeat the above steps iteratively until the preset iteration stop condition is met, indicating that the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1200 times.

[0214] S600: Use the trained model to process new sample data to evaluate the digital management maturity level it belongs to.

[0215] Specifically, use the trained model to process new sample data to evaluate the digital management maturity level it belongs to. In one embodiment, there is a new sample data set of an enterprise. The new samples are input into the trained feature extraction model for feature extraction. Further, the data after feature extraction is input into the trained data dimensionality reduction model for data dimensionality reduction. Further, the dimensionality-reduced features are input into the trained classifier model, and the digital management maturity level of the enterprise is classified according to the features. In this embodiment, the classification categories include: "high maturity level", "medium maturity level", and "low maturity level", a total of 3 categories. Based on this, the process of model training and application of the present invention is as follows Figure 3 as shown.

[0216] Adopt a generative adversarial network algorithm based on a momentum optimization variant for data augmentation. Through the collaborative work of the generator and the discriminator, realistic enterprise management data is effectively generated, which can overcome the problem of insufficient training caused by limited data samples, improve the generalization ability of the model, enhance the diversity and authenticity of the generated data, overcome the problem of model overfitting caused by insufficient data volume, and improve the prediction accuracy of the model.

[0217] Adopt a circular dynamic optimization algorithm to optimize the parameters of the neural network, simulate different stages in the cell cycle, and enhance the adaptability and accuracy of the model to complex data environments by adjusting the optimization strategy stage by stage, effectively reducing the risk of overfitting, and being able to effectively jump out of the local optimal solution, thereby improving the accuracy and efficiency of feature extraction.

[0218] Adopt an autoencoder based on lightweight fuzzy weights to perform dimensionality reduction processing on enterprise management data. By dynamically adjusting the fuzzy logic weights in the weight update process, the accuracy of the dimensionality-reduced data and the generalization ability of the network are improved. By adjusting the fuzzy logic weights, the autoencoder not only reduces the reconstruction error, but also can more accurately reflect the structural characteristics of the original data, enhancing the stability and reliability of data processing.

[0219] The limit learning machine algorithm based on fractional gradient adjustment is adopted to improve the classifier training process. Through the meticulous control of fractional derivatives, the processing ability of nonlinear problems and the generalization of the classifier are optimized. The fractional gradient adjustment enables the limit learning machine to more finely adjust the weights during the learning process, reduces the risk of model overfitting, and improves the ability to handle complex classification problems.

[0220] The above-disclosed are only one or more preferred embodiments of the present application, and the scope of rights of the present application cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present application still fall within the scope covered by the present application.

Claims

1. An artificial intelligence-based method for evaluating the maturity of enterprise digital management, characterized in that, It includes the following steps: Collect enterprise management data and perform annotation; The sample generation is carried out by using the generative adversarial network algorithm based on the momentum optimization variant to obtain the enterprise management data augmentation; the generative adversarial network algorithm based on the momentum optimization variant includes a generator and a discriminator, and the Adam optimizer with a dynamic learning rate is used to optimize the parameters of the generator and the discriminator. The optimization method is as follows: In the formula, θ bG and θ bD respectively represent the parameters of the generator and the discriminator; η bt is the learning rate for the t-th iteration; and are the first and second moment estimates; ∈ is a small positive number added to prevent division by zero; The quality of enterprise management data optimized by the adaptive feature enhancement technology is enhanced by dynamically adjusting the key features in the generated enterprise management data according to the feedback of the discriminator, so as to enhance the adaptability and accuracy of the model to complex enterprise management scenarios. Among them, the feature difference evaluation function F diff () is a function for quantifying the key feature differences between the generated enterprise management data and the real enterprise management data, and the calculation method is as follows: In the formula, x gen and x real represent the feature vectors of the generated enterprise management data and the real enterprise management data respectively; Kr is the number of features; ω k is the importance weight of the k-th feature; Input the augmented enterprise management data into the feature extraction model for training the feature extraction model; Input the enterprise management data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model. The steps include: Initialize all weights of the autoencoder; In the encoder training stage, input the enterprise management data after feature extraction and transform it into a low-dimensional feature representation. Each input enterprise management data is gradually compressed through the encoder hierarchy. Each layer of the encoder calculates a new output through the output of the previous layer and the weights of the current layer. Each layer of the encoder calculates a new output through the output of the previous layer and the weights of the current layer. The steps also include: The output h of each layer of the encoder pj is calculated as follows: where h pj is the output of the j-th layer; x pi is the input vector or the output of the previous layer; w pij is the weight connecting the i-th input and the j-th output; b pj is the bias of the j-th layer; Re p ( ) is the ReLU activation function; After the features are encoded, adjust the network weights according to the inherent ambiguity of the enterprise management data. Evaluate the importance and contribution degree of each feature through fuzzy logic, and dynamically adjust the weights according to the evaluation results; The decoder receives the low-dimensional features output by the encoder and attempts to reconstruct the original input data. Each low-dimensional feature is expanded layer by layer until data with the same dimension as the original input is reconstructed; Calculate the error by comparing the difference between the reconstructed data and the original input data, and update the weights in the network through the backpropagation algorithm to reduce the reconstruction error; Adopt an adaptive feature attenuation strategy to optimize the dimensionality reduction quality and reconstruction accuracy of enterprise management data by dynamically adjusting the retention or attenuation degree of features in the decoding stage during the training of the feature dimensionality reduction model of the autoencoder; Repeat the above steps iteratively until the preset stop iteration condition is met, which means the model training is completed; Input the dimensionality-reduced enterprise management data into the classifier model for training the classifier; Use the trained model to process new sample data to evaluate the digital management maturity of the enterprise it belongs to.

2. The method for evaluating the maturity of enterprise digital management based on artificial intelligence according to claim 1, wherein Adopt a generative adversarial network algorithm based on a momentum optimization variant to generate samples to obtain enterprise management data augmentation. The steps also include: Among them, the goal of the generator is to generate realistic enterprise management data to deceive the discriminator; The task of the discriminator is to distinguish between real enterprise management data and fake enterprise management data generated by the generator.

3. The method for evaluating the maturity of enterprise digital management based on artificial intelligence according to claim 2, wherein, The training process of the generative adversarial network algorithm based on the momentum optimization variant includes: Initialize the parameters of the generative adversarial network model based on the momentum optimization variant. Among them, the initialization method adopts the Gaussian distribution strategy for initialization, expressed as: In the formula, represents the initial weight of the generator; represents the initial weight of the discriminator; n bin and n bou are the number of nodes in the input layer and the output layer respectively; Z TF () is the Gaussian distribution function; In each training cycle, the generator generates a batch of fake enterprise management data, and the discriminator evaluates the difference between these enterprise management data and the real enterprise management data. The feedback of the discriminator is used to adjust the parameters of the generator to make the enterprise management data generated by it more realistic. Among them, the ways of generating enterprise management data and evaluation are expressed as: D out = f bD (G bG (z; θ bG )); θ bD ), Lb G＝ -log(D out ) - log(Θ afe (x gen ))), L bD = -[y log(D out ) + (1 - y) log(1 - D out )] In the formula, D out is the output of the discriminator, Θ afe (x gen ) is the generated enterprise management data after feature enhancement, representing the probability of being judged as real enterprise management data; G bG (z; θ bG ) is the enterprise management data generated by the generator according to the input z and the parameter θ bG ; f bD () is the function of the discriminator; L bG and L bD are the loss functions of the generator and the discriminator respectively; y is the label, 1 for real enterprise management data and 0 for generated enterprise management data; Based on the feature difference evaluation, the calculation method of the adaptive feature enhancement strategy function Θ afe () is expressed as: In the formula, λ gr is the learning rate; represents the gradient of the feature difference evaluation function for the generated enterprise management data x gen ; the k-th feature of the feature vector of the generated enterprise management data, the k-th feature of the feature vector of the real enterprise management data; According to the feedback of the discriminator, continuously adjust the strategy for generating enterprise management data to adapt to the changes in the distribution of enterprise management data. The adjustment of the generation strategy is achieved by means of an adaptive mechanism, which is expressed as: In the formula, γ bG is the dynamic adjustment factor of the generator learning rate; α bG is the learning step size; Δθ bG is the adjustment amount of the parameter, and the calculation method of the learning rate dynamic adjustment factor γ bG is expressed as: In the formula, δ bG is the attenuation rate of the adjustment factor; is the adjustment amount of the parameter θ bG at the i-th iteration; represents the two-norm of the adjustment amount; After a fixed period, the generation effect of the model is evaluated, and the parameters of the generator and discriminator are updated accordingly. The way to update the parameters is expressed as: In the formula, β bG is the first momentum factor; β bD is the second momentum factor; is the updated generator parameter; is the generator parameter before update; is the updated discriminator parameter; is the discriminator parameter before update; Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed.

4. The method for evaluating the maturity of enterprise digital management based on artificial intelligence according to claim 1, wherein Input the augmented enterprise management data into the feature extraction model for training the feature extraction model. The steps also include: Initialize the neural network model; In the growth period, quickly adapt to the characteristics of enterprise management data by accelerating the fine-tuning of weights; In the synthesis period, perform weight synthesis; In the evaluation period, evaluate the current network performance and decide whether to adjust the cycle optimization strategy. If so, return to the growth period. If not, enter the pre-splitting stage; In the pre-splitting stage, simulate the preparation stage of cell division and make fine adjustments to the weights; In the splitting stage, implement structural adjustment; In the dormancy period, adjust the learning rate and perform learning rate decay; Repeat the above steps until the preset iteration stop condition is met, indicating that the model training is completed.

5. The method for evaluating the maturity of enterprise digital management based on artificial intelligence according to claim 1, characterized in that Input the dimensionality-reduced enterprise management data into the classifier model for training the classifier, and the steps further include: In the initialization stage, the weights and biases from the input layer to the hidden layer are randomly initialized; Calculate the representation of each sample in the hidden layer; Use the least squares method to adjust the output layer weights and minimize the difference between the actual output and the expected output; Adopt fractional gradient adjustment to optimize the weights by calculating the fractional-order derivative of the gradient; Repeat the above steps until the preset iteration stop condition is met, indicating that the model training is completed.

6. The method for evaluating the maturity of enterprise digital management based on artificial intelligence according to claim 1, characterized in that After the features are encoded, adjust the network weights according to the inherent ambiguity of the enterprise management data, evaluate the importance and contribution degree of each feature through fuzzy logic, and dynamically adjust the weights according to the evaluation results. The steps further include: The adjustment method is expressed as: δw pij = μ p (x pi , h pj )·(x pi - h pj ) In the formula, is the adjusted weight parameter; δw pij is the adjustment amount of the weight; μ p (x pi , h pj ) is the weight adjustment factor based on fuzzy logic; α p is the learning rate, and the weight adjustment factor μ p (x pi , h pj ) is dynamically calculated according to the relationship between the input x pi and the output h pj , and the calculation method is expressed as: In the formula, x pi is the data input to the i-th layer of the autoencoder; h pj is the output of the i-th layer of the autoencoder; σ pμ is the fuzzy factor.

7. The method for evaluating the maturity of enterprise digital management based on artificial intelligence according to claim 1, wherein The decoder receives the low-dimensional features output by the encoder and attempts to reconstruct the original input data. Each low-dimensional feature is expanded layer by layer until data with the same dimension as the original input is reconstructed. The steps further include: the output y of the decoder pj is calculated as follows: In the formula, y pj is the reconstructed output; w pji is the weight in the decoder; b pi is the bias in the decoder.

8. The method for evaluating the maturity of enterprise digital management based on artificial intelligence according to claim 1, wherein By comparing the differences between the reconstructed data and the original input data, calculating the error, and updating the weights in the network through the backpropagation algorithm to reduce the reconstruction error, the steps further include: the error E p and the update manner of the gradient of the weights is expressed as: Where, E p is the reconstruction error; x pj is the original input data; y pj is the output of the decoder; Re′ p is the derivative of the ReLU activation function, and the calculation of the gradient is expressed as: Where, z pj is the input weighted sum of the j-th neuron in the decoder; h pi is the output of the encoder, that is, the input of the decoder; Re′ p (z pj ) is the derivative of the activation function at z pj .

Images