Production material demand prediction method and system based on artificial intelligence

Abstract

The invention discloses a production material demand prediction method and system based on artificial intelligence, and belongs to the technical field of demand prediction, and the method comprises the steps: production material data integration, production material data optimization, construction of a production material demand prediction model, and generation of a prediction report. According to the scheme, feature vectors are clustered, feature subsets are constructed, the local classification confidence coefficient of data is obtained, the local classification confidence coefficient is combined with a weight regularization item to construct a loss function, the weight of the feature vectors is updated, feature importance is analyzed to obtain a smooth threshold vector, feature vector screening is carried out, and the complexity of the data is reduced; node attribute similarity and neighbor difference quantization are combined, bimodal dynamic weight is calculated, node attributes are updated, a second-order disturbance term and deviation quadratic sum are introduced, the node attributes are mapped, noise distribution is dynamically corrected through gradient, potential variables are sampled, and node attributes are reconstructed by introducing a cubic deviation term. And the accuracy and timeliness of production material demand prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of demand forecasting, and specifically refers to a production material demand forecasting method and system based on artificial intelligence. Background Art

[0002] The production material demand forecasting method is an intelligent solution that uses artificial intelligence and big data analysis technologies to predict the future production material demand by integrating multi-dimensional data of historical production materials, realizes the precise control of material inventory, reduces warehousing costs and out-of-stock risks, improves the reliability of production plans and the collaborative efficiency of the supply chain, and helps enterprises achieve lean production.

[0003] However, in the existing production material demand forecasting methods, there are problems that the production material demand forecasting involves a large number of features, many of which are highly correlated or contain noise, and the factors that truly affect the demand may be masked by a large number of redundant features, resulting in low accuracy and efficiency of demand forecasting; in the existing production material demand forecasting methods, there are problems that the production material demand is affected by the interaction of many factors, making the material demand exhibit highly non-linear and dynamic change attributes, and traditional methods are difficult to accurately capture these attributes, resulting in large demand forecasting deviations and low timeliness. Summary of the Invention

[0004] In view of the above situation, to overcome the defects of the prior art, the present invention provides a production material demand prediction method and system based on artificial intelligence. In the existing production material demand prediction methods, a large number of features are involved in the production material demand prediction, and many of these features are highly correlated or contain noise. The factors that truly affect the demand may be masked by a large number of redundant features, resulting in low accuracy and efficiency of demand prediction. In this solution, the feature vectors are clustered and feature subsets are constructed to obtain the local classification confidence of the data. A loss function is constructed by combining with the weight regularization term, the weights of the feature vectors are updated, and the normalized feature weight vectors are obtained by combination. The importance of the features is analyzed to obtain the smoothing threshold vector. Feature vector screening is performed based on the two vectors to complete the optimization, which can focus more on the factors that truly affect the demand, reduce the influence of noise or redundant features, simplify the feature space, reduce the complexity of the data, and improve the accuracy and efficiency of production material demand prediction. In the existing production material demand prediction methods, the production material demand is affected by the interaction of many factors, making the material demand exhibit highly nonlinear and dynamic changing attributes. Traditional methods are difficult to accurately capture these attributes, resulting in large demand prediction deviations and low timeliness. In this solution, the node attribute similarity and neighbor difference quantification are combined, the bimodal dynamic weights are calculated and the node attributes are updated. The second-order perturbation term and the sum of squared deviations are introduced to map the node attributes. The noise distribution is corrected dynamically by the gradient, the latent variables are sampled, and the cubic deviation term is introduced to reconstruct the node attributes. After multiple iterations, the final node attributes are obtained and predictions are made, improving the accuracy of the description of the material relationship, enabling the updated node attributes to better reflect the actual state of the material under complex relationships, and enabling the latent variables to better reflect the internal characteristics of the material demand affected by multiple factors, thereby improving the accuracy and timeliness of production material demand prediction.

[0005] The technical solution adopted by the present invention is as follows: The production material demand prediction method based on artificial intelligence provided by the present invention includes the following steps:

[0006] Step S1: Integrate production material data;

[0007] Step S2: Optimize production material data;

[0008] Step S3: Construct a production material demand prediction model;

[0009] Step S4: Generate a prediction report.

[0010] Further, in step S1, the integration of production material data is to collect historical production material time series data and process historical production material time series data;

[0011] The collection of historical production material time series data is to collect timestamps, material data, production data, order data, and demand quantities;

[0012] The processed historical production material time series data is obtained by performing data cleaning, data normalization, data encoding, and dataset construction on the historical production material time series data, resulting in a production material time series dataset.

[0013] Further, in step S2, the optimization of the production material data specifically includes the following steps:

[0014] Step S21: Feature vector clustering; using the K-means clustering algorithm, all feature vectors in the production material time series dataset E are divided into K feature clusters, and the weight of each feature vector is initialized to 1, generating an initial feature weight vector of Q dimensions;

[0015] Step S22: Constructing a feature subset; creating A empty sets, for each feature cluster, randomly select feature vectors from it, and sequentially assign them to A sets, and remove the assigned feature vectors from the feature cluster; if the number of feature vectors in the feature cluster is not zero after removal, evenly and randomly assign the remaining feature vectors in the feature cluster to any one of the sets, and remove the assigned feature vectors from the feature cluster again until the number of feature vectors in each feature cluster is zero, obtaining A feature subsets; where, Z k is the number of feature vectors in the k-th feature cluster, k is the feature cluster index, is the floor operator;

[0016] Step S23: Updating the feature vector weight; including the following steps:

[0017] Step S231: Calculating the local classification confidence; based on the feature vector weight, calculate the weighted distance between every two data in the production material time series dataset under each feature subset, then introduce a temperature coefficient to control the distribution smoothness, obtain the similarity between every two data, and perform normalization to obtain the local classification confidence of each data under each feature subset;

[0018] Step S232: Defining a loss function; combining the local classification confidence with a weight regularization term to construct a loss function;

[0019] Step S233: Updating; preset a first convergence threshold P1 and a maximum number of updates C1, for each feature subset, use the gradient ascent method to iteratively update the feature vector weight until or when the maximum number of iterations is reached, stop updating and output the feature weight vector at this time; where, ω a is the feature weight vector corresponding to the a-th feature subset, is ω a corresponding loss function, J maxis the maximum loss function value obtained during the current update process;

[0020] Step S234: Combination; Recombine the updated feature weight vectors of A feature subsets according to the indexes of the initial feature weight vectors to obtain a Q-dimensional feature weight vector and perform linear normalization to obtain a Q-dimensional normalized feature weight vector h;

[0021] Step S24: Feature importance analysis; Calculate the Pearson correlation coefficient between every two feature vectors in the production material time series dataset to obtain the redundancy score of each feature vector with the remaining feature vectors. Combine the normalized feature weight vector and the redundancy score to generate a deterministic threshold vector, and perform a moving average filter on the deterministic threshold vector to obtain a smoothed threshold vector;

[0022] Step S25: Feature vector screening; If the value of the v-th column of the normalized feature weight vector h is greater than the value of the v-th column of the smoothed threshold vector, then use the v-th column feature vector in the production material time series dataset as a valid feature vector; otherwise, use the v-th column feature vector as a redundant feature vector; Construct a demand prediction dataset based on all valid feature vectors to complete the optimization.

[0023] Further, in step S3, the construction of the production material demand prediction model specifically includes the following steps:

[0024] Step S31: Construct an adjacency matrix; Use the demand prediction dataset as input data, regard each data in the demand prediction dataset as a node, and the node attribute is the corresponding production material time series data. Preset a similarity threshold L. If the cosine similarity between two node attributes is less than the similarity threshold, there is an edge between the corresponding two nodes, and the weight of the edge is 1; otherwise, there is no edge between the corresponding two nodes, and the weight of the edge is 0; Obtain the adjacency matrix and perform symmetric normalization processing on the adjacency matrix;

[0025] Step S32: Update node attributes; Combine the node attribute similarity and neighbor difference quantization, calculate the bimodal dynamic weight between each node and its adjacent nodes, and complete the update of the node attributes based on the bimodal dynamic weight; The formula used is as follows:

[0026] ;

[0027] ;

[0028] where p1, p2, and p3 are node indexes, is the updated node attribute of node p1, s p1 , s p2 and s p3They are the original node attributes of node p1, node p2, and node p3 respectively. is the set of adjacent nodes of node p1. is the bimodal dynamic weight between node p1 and node p2. W and q are learnable weight matrices and weight vectors respectively, and λ is a control factor. and are the node attribute similarity and neighbor difference quantization respectively. is the activation function. is the concatenation operation. T is the transpose operation, and δ is a regulation factor. is the square of the L2 norm.

[0029] Step S33: Latent variable sampling; By introducing a second-order perturbation term and the sum of squared deviations, the node attributes are mapped to the mean and variance in the latent space, and the noise distribution is dynamically corrected through the gradient of the KL divergence loss, and the latent variables are sampled based on the corrected distribution. The formula used is as follows:

[0030] ;

[0031] ;

[0032] ;

[0033] In the formula, x p1 is the latent variable corresponding to node p1, μ p1 and are the mean and variance of the attribute of node p1 mapped to the latent space respectively. and are two independent feedforward neural networks. N att is the total dimension of the node attributes, d is the dimension index, z d is the global attribute mean in the d-th dimension. , is the value of the node attribute of node p1 in the d-th dimension. is the small perturbation of the node attribute of node p1 in the d-th dimension. and are the second-order perturbation term and the sum of squared deviations respectively. is the element-wise multiplication operator. is a random noise vector following the standard normal distribution. , N node is the number of nodes, e is the node index. is the standard normal distribution with a mean of 0 and a covariance matrix of the identity matrix I. is the KL divergence loss. is the e-th component of the noise vector. is The partial derivative of with respect to is the partial derivative of the network r μ with respect to ;

[0034] Step S34: Reconstruct node attributes; introduce a cubic bias term and use latent variables to reconstruct node attributes; the formula used is as follows:

[0035] ;

[0036] where, is the reconstructed node attribute of node p1, is the decoder that uses latent variables to reconstruct node attributes, is the mean of the latent variables in the w-th dimension, , is the value of the latent variable corresponding to node p1 in the w-th dimension, N var is the total dimension of the latent variables, w is the dimension index, is the cubic bias term;

[0037] Step S35: Prediction; preset a second convergence threshold P2 and a maximum number of iterations C2, repeat steps S32 - S34 until the mean square error between the reconstructed node attributes and the original node attributes is less than or equal to P2 or the maximum number of iterations is reached, then stop the iteration. Input the node attributes U at this time into the spatio-temporal convolutional layer to capture spatio-temporal dependencies, and then perform mean pooling to obtain the final node attributes U final , and predict the production material demand through a fully connected layer, outputting the predicted data label.

[0038] Furthermore, in step S4, the generation of the prediction report is to collect and process real-time production material time series data, input it into the production material demand prediction model for prediction, obtain the demand for the corresponding production material according to the output data label, and generate a production material demand prediction report;

[0039] The collection of real-time production material time series data is to collect timestamps, material data, production data, and order data;

[0040] The processing of real-time production material time series data is to perform data cleaning, data normalization, and data encoding on the real-time production material time series data.

[0041] The production material demand prediction system based on artificial intelligence provided by the present invention includes a production material data integration module, a production material data optimization module, a module for constructing a production material demand prediction model, and a module for generating a prediction report;

[0042] The production material data integration module collects and processes historical production material time series data and sends the data to the production material data optimization module;

[0043] The production material data optimization module clusters the feature vectors and constructs feature subsets to obtain the local classification confidence of the data, combines with the weight regularization term to construct a loss function, updates the feature vector weights, combines to obtain a normalized feature weight vector, analyzes the feature importance to obtain a smoothing threshold vector, and performs feature vector screening based on the two vectors to complete the optimization, and sends the data to the module for constructing the production material demand prediction model;

[0044] The module for constructing the production material demand prediction model combines the node attribute similarity and neighbor difference quantization, calculates the bimodal dynamic weight and updates the node attributes, introduces the second-order perturbation term and the sum of squared deviations, maps the node attributes, dynamically corrects the noise distribution through the gradient, samples the latent variables, introduces the cubic deviation term to reconstruct the node attributes, iterates multiple times to obtain the final node attributes and makes predictions, and sends the data to the module for generating the prediction report;

[0045] The module for generating the prediction report obtains the demand for the corresponding production material according to the data label output by the production material demand prediction model and generates a production material demand prediction report.

[0046] The beneficial effects achieved by the present invention using the above solution are as follows:

[0047] (1) Aiming at the problem in the existing production material demand prediction method that there are a large number of features involved in the production material demand prediction, many of which are highly correlated or contain noise, and the factors that really affect the demand may be covered by a large number of redundant features, resulting in low accuracy and efficiency of the demand prediction. This solution clusters the feature vectors and constructs feature subsets to obtain the local classification confidence of the data, combines with the weight regularization term to construct a loss function, updates the feature vector weights, combines to obtain a normalized feature weight vector, analyzes the feature importance to obtain a smoothing threshold vector, and performs feature vector screening based on the two vectors to complete the optimization, which can focus more on the factors that really affect the demand, reduce the influence of noise or redundant features, simplify the feature space, reduce the complexity of the data, and improve the accuracy and efficiency of the production material demand prediction.

[0048] (2)In view of the problem that in the existing production material demand prediction methods, the production material demand is affected by the interaction of many factors, making the material demand exhibit highly non-linear and dynamic changing properties, and traditional methods are difficult to accurately capture these properties, resulting in large demand prediction deviations and low timeliness, this solution combines the node attribute similarity and the quantification of neighbor differences, calculates the bimodal dynamic weight and updates the node attributes, introduces the second-order perturbation term and the sum of squared deviations, maps the node attributes, dynamically corrects the noise distribution through the gradient, samples the latent variables, introduces the cubic deviation term to reconstruct the node attributes, and obtains the final node attributes through multiple iterations for prediction, improving the accuracy of the description of the material relationship, enabling the updated node attributes to better reflect the actual state of the material under complex relationships, mapping the node attributes to the latent space more accurately, enabling the latent variables to better reflect the internal characteristics of the material demand affected by multiple factors, and making more precise use of the material relationship information, thereby improving the accuracy and timeliness of the production material demand prediction. Description of the Drawings

[0049] Figure 1 It is a schematic flowchart of the production material demand prediction method based on artificial intelligence provided by the present invention;

[0050] Figure 2 It is a schematic diagram of the production material demand prediction system based on artificial intelligence provided by the present invention;

[0051] Figure 3 It is a schematic flowchart of step S2;

[0052] Figure 4 It is a schematic flowchart of step S3.

[0053] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. Detailed Embodiments

[0054] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0055] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.

[0056] Example 1. Refer to Figure 1 , the production material demand prediction method based on artificial intelligence provided by the present invention includes the following steps:

[0057] Step S1: Integrate production material data; collect and process historical production material time series data;

[0058] Step S2: Optimize production material data; cluster the feature vectors and construct a feature subset to obtain the local classification confidence of the data, combine with the weight regularization term to construct a loss function, update the feature vector weights, combine to obtain a normalized feature weight vector, analyze the feature importance to obtain a smoothing threshold vector, and perform feature vector screening based on the two vectors to complete the optimization;

[0059] Step S3: Construct a production material demand prediction model; combine the node attribute similarity and neighbor difference quantification, calculate the bimodal dynamic weight and update the node attributes, introduce a second-order perturbation term and the sum of squared deviations, map the node attributes, dynamically correct the noise distribution through the gradient, sample the latent variables, introduce a cubic deviation term to reconstruct the node attributes, and obtain the final node attributes through multiple iterations for prediction;

[0060] Step S4: Generate a prediction report; obtain the demand for the corresponding production material according to the data label output by the production material demand prediction model, and generate a production material demand prediction report.

[0061] Example 2. Refer to Figure 1 , based on the above example, in step S1, integrating production material data is to collect historical production material time series data and process historical production material time series data;

[0062] Collecting the historical production material time series data is to collect the time stamp, material data, production data, order data, and demand;

[0063] The material data includes the material type, inventory quantity, inventory turnover rate, and daily consumption;

[0064] The production data includes the daily production volume, production batch plan, production cycle, production scheduling, and production equipment status;

[0065] The order data includes the order quantity and the order product type;

[0066] The processed historical production material time series data is obtained by performing data cleaning, data normalization, data encoding, and dataset construction on the historical production material time series data, resulting in a production material time series dataset;

[0067] The data cleaning is to remove the error values, missing values, and outliers in the data;

[0068] The data normalization is to unify the numerical data into the same range using the maximum-minimum scaling method;

[0069] The data encoding is to convert the categorical data into numerical data using One-Hot encoding;

[0070] The dataset construction is to use the material demand quantity as the data label, and then construct a production material time series dataset based on the historical production material time series data after data cleaning, data normalization, and data encoding.

[0071] Example 3, refer to Figure 1 and Figure 3 , based on the above example, in step S2, the optimization of the production material data specifically includes the following:

[0072] Step S21: Feature vector clustering; in the production material demand prediction, the original data is usually high-dimensional and redundant, and direct use will lead to high computational complexity and model overfitting. By using K-means clustering to group similar features into the same cluster, the complexity of the feature space is reduced; using the K-means clustering algorithm, all feature vectors in the production material time series dataset E are divided into K feature clusters, the elbow method is used to determine the value of K, and the weight of each feature vector is initialized to 1, generating an initial feature weight vector of Q dimensions;

[0073] Step S22: Constructing feature subsets; a single feature subset may miss key features or introduce noise, while the full set of features has a high computational cost. By constructing feature subsets, it is ensured that each subset covers different feature combinations and key information loss is avoided; create A empty sets, , for each feature cluster, randomly select feature vectors from it, and assign them to the A sets in turn, and remove the assigned feature vectors from the feature cluster; if the number of feature vectors in the feature cluster is not zero after removal, evenly and randomly assign the remaining feature vectors in the feature cluster to any one of the sets, and remove the assigned feature vectors from the feature cluster again until the number of feature vectors in each feature cluster is zero, obtaining A feature subsets; where Z k is the number of feature vectors in the kth feature cluster, and k is the feature cluster index. is the floor operator;

[0074] Step S23: Update the eigenvector weights; including the following steps:

[0075] Step S231: Calculate the local classification confidence; the demand patterns of similar materials in production data may be misjudged due to noise. The stable demand patterns are highlighted by the local classification confidence, and the interference of outliers is suppressed; based on the eigenvector weights, calculate the weighted distance between every two data in the production material time series dataset under each feature subset, and then introduce a temperature coefficient to control the distribution smoothness to obtain the similarity between every two data, and perform normalization to obtain the local classification confidence of each data under each feature subset; the formula used is as follows:

[0076] ;

[0077] ;

[0078] ;

[0079] In the formula, is the local classification confidence of data b in the production material time series dataset under the a-th feature subset i ; and are respectively the weighted distance and similarity between data b i and data b j under the a-th feature subset; is the similarity between data b n and data b j under the a-th feature subset; b i 、b j and b n are respectively the i-th, j-th, and n-th data in the production material time series dataset, i, j, and n are data indices, a is the feature subset index, and are respectively the nearest neighbor sets of data b i and data b n found based on the weighted distance, and the number of data in the nearest neighbor set is equal to , H is the number of data in the production material time series dataset, m is the eigenvector index, M a is the number of eigenvectors in the a-th feature subset, is the square value of the m-th eigenvector weight in the a-th feature subset, and are respectively data b in the a-th feature subseti and data b j The value on the m-th eigenvector, is the temperature coefficient, ; is the indicator function. When y i =y j at that time, is 1, otherwise, is 0; y i and y j are the data labels corresponding to data b i and data b j respectively;

[0080] Step S232: Define the loss function; Feature weights may overly rely on a few high-variance features and ignore long-term stable features. By defining the loss function, prevent the weights from overly concentrating on a single feature and prevent overfitting; Combine the local classification confidence with the weight regularization term to construct the loss function; The formula used is as follows:

[0081] ;

[0082] In the formula, ω a is the eigenvector of feature weights corresponding to the a-th feature subset, H is the number of data in the production material time series dataset, is the loss function corresponding to ω a , is the weight regularization term, γ is the regularization adjustment factor, γ = 0.01;

[0083] Step S233: Update; Dynamically adjust the weights by the gradient ascent method to ensure efficiency and accuracy; Preset the first convergence threshold P1 and the maximum number of updates C1. For each feature subset, use the gradient ascent method to iteratively update the eigenvector weights until or reach the maximum number of iterations and then stop the update, and output the eigenvector of feature weights at this time; The formula used is as follows:

[0084] ;

[0085] In the formula, and are the eigenvectors of feature weights corresponding to the a-th feature subset at the c-th update respectively, η is the learning rate, η = 0.01, c is the update count index, is corresponding loss function, is for gradient of, J max is the maximum loss function value obtained during the current update process, P1 = 10 -4 , C1 = 100;

[0086] Step S234: Combination; Recombine the updated feature weight vectors of A feature subsets according to the index of the initial feature weight vector to obtain a Q-dimensional feature weight vector and perform linear normalization to obtain a Q-dimensional normalized feature weight vector h;

[0087] Step S24: Feature importance analysis; Multicollinearity often exists in the production material time series data. Simply relying on weights may retain redundant features, resulting in poor model interpretability. By feature importance analysis, weight mutations are eliminated, robustness is improved, key features are ensured to be retained first, and interference from duplicate features is avoided; To further eliminate redundant features, calculate the Pearson correlation coefficient between every two feature vectors in the production material time series dataset to obtain the redundancy score of each feature vector with the remaining feature vectors. Combine the normalized feature weight vector and the redundancy score to generate a deterministic threshold vector, and perform a moving average filter on the deterministic threshold vector to obtain a smoothed threshold vector; The formulas used are as follows:

[0088] ;

[0089] ;

[0090] ;

[0091] In the formula, is the Pearson correlation coefficient between the u-th and v-th feature vectors, R v is the redundancy score corresponding to the v-th feature vector, u and v are feature vector indices, max(R) is the maximum redundancy score, α is a weight adjustment factor, α = 0.7, h v is the weight value corresponding to the v-th feature vector in the normalized feature weight vector h, f v and are the deterministic threshold and the smoothed threshold corresponding to the v-th feature vector respectively, o is a temporary traversal index, is the deterministic threshold corresponding to the v+o-th feature vector;

[0092] Step S25: Feature vector screening; Improve the quality of training data through feature vector screening to make the model more focused on key factors; If the value of the v-th column of the normalized feature weight vector h is greater than the value of the v-th column of the smoothed threshold vector, then use the v-th column feature vector in the production material time series dataset as a valid feature vector; Otherwise, use the v-th column feature vector as a redundant feature vector; Construct a demand prediction dataset based on all valid feature vectors to complete the optimization.

[0093] By performing the above operations, for the problem in the existing production material demand forecasting method that there are a large number of features involved in the production material demand forecasting, many of which are highly correlated or contain noise, and the factors that truly affect the demand may be masked by a large number of redundant features, resulting in low accuracy and efficiency of demand forecasting. In this solution, the feature vectors are clustered and feature subsets are constructed to obtain the local classification confidence of the data. A loss function is constructed in combination with the weight regularization term, the weights of the feature vectors are updated, the normalized feature weight vectors are combined, the feature importance is analyzed to obtain the smoothing threshold vector, and the feature vectors are screened based on the two vectors to complete the optimization, which can focus more on the factors that truly affect the demand, reduce the influence of noise or redundant features, simplify the feature space, reduce the complexity of the data, and improve the accuracy and efficiency of production material demand forecasting.

[0094] Example 4, refer to Figure 1 and Figure 4 , based on the above example, in step S3, constructing the production material demand forecasting model specifically includes the following steps:

[0095] Step S31: Construct an adjacency matrix; in production material demand forecasting, there may be implicit correlations in the time series data of different materials, but traditional methods are difficult to quantify this relationship. Transforming the non-linear relationship of material demand data into a graph structure can avoid the interference of noise edges, and symmetric normalization can balance the node degree differences and prevent high-demand materials from overly dominating model learning. Taking the demand forecasting data set as the input data, each data in the demand forecasting data set is used as a node, and the node attribute is the corresponding production material time series data. A similarity threshold L is preset. If the cosine similarity between two node attributes is less than the similarity threshold, there is an edge between the corresponding two nodes, and the weight of the edge is 1; otherwise, there is no edge between the corresponding two nodes, and the weight of the edge is 0. An adjacency matrix is obtained, and the adjacency matrix is symmetrically normalized. The formula used is as follows:

[0096] ;

[0097] In the formula, F and are the original adjacency matrix and the symmetrically normalized adjacency matrix respectively, G is the degree matrix, I is the identity matrix, and L = 0.8;

[0098] Step S32: Update the node attributes; material demand is affected by multiple factors, and a single similarity measure cannot take into account both semantic associations and local structural differences. The bimodal dynamic weight simultaneously fuses the node attribute similarity and neighbor difference quantification, which can not only identify collaborative demand patterns but also perceive local anomalies. Combining the node attribute similarity and neighbor difference quantification, calculate the bimodal dynamic weight between each node and its adjacent nodes, and complete the update of the node attributes based on the bimodal dynamic weight. The formula used is as follows:

[0099] ;

[0100] ;

[0101] Among them, p1, p2, and p3 are node indices, is the node attribute after the update of node p1, s p1 , s p2 and s p3 are the original node attributes of node p1, node p2, and node p3 respectively, is the set of adjacent nodes of node p1, is the bimodal dynamic weight between node p1 and node p2, W and q are learnable weight matrices and weight vectors respectively, λ is a control factor, λ = 0.1, and are the node attribute similarity and neighbor difference quantization respectively, is the activation function, is the concatenation operation, T is the transpose operation, δ is a regulation factor, δ = 0.3, is the square of the L2 norm;

[0102] Step S33: Sampling of latent variables; There is uncertainty in the material requirement data, and traditional deterministic coding is difficult to represent its probability distribution. The propagation error of the input noise to the mean is quantified through the second-order perturbation term to improve the robustness to data fluctuations. The sum of squared deviations constrains the dispersion of the latent space to prevent the latent variables from deviating too much from the global mean and ensure the generation of reasonable demand scenarios. The KL divergence gradient corrects the noise to dynamically adjust the sampling direction and alleviates the posterior collapse problem under sparse demand data; By introducing the second-order perturbation term and the sum of squared deviations, the node attributes are mapped to the mean and variance of the latent space, and the noise distribution is dynamically corrected through the gradient of the KL divergence loss, and the latent variables are sampled based on the corrected distribution; The formulas used are as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] In the formula, x p1 is the latent variable corresponding to node p1, μ p1 and are the mean and variance of the attribute of node p1 mapped to the latent space respectively, and are two independent feedforward neural networks, N att is the total dimension of the node attributes, d is the dimension index, zd is the global attribute mean in the d-th dimension, , is the value of the node attribute of node p1 in the d-th dimension, is the small perturbation of the node attribute of node p1 in the d-th dimension, and are the second-order perturbation term and the sum of squared deviations respectively, is the element-wise multiplication operator, is a random noise vector following the standard normal distribution, , N node is the number of nodes, e is the node index, is the standard normal distribution with mean 0 and covariance matrix being the identity matrix I, is the KL divergence loss, is the e-th component of the noise vector, is the partial derivative of with respect to is the network r μ the partial derivative of with respect to

[0107] Step S34: Reconstruct the node attributes; The material requirement reconstruction needs to balance the overall trend fitting and local mutation recovery. The cubic deviation term amplifies the penalty for the potential variable deviating from the mean, forcing the model to pay more attention to abnormal demand patterns, which is superior to the over-sensitivity of the traditional L2 loss to small fluctuations. The decoder works in cooperation with the cubic term to enhance the generation ability for outliers while ensuring the reconstruction accuracy of the mainstream demand; Introduce the cubic deviation term and use the potential variable to reconstruct the node attributes; The formula used is as follows:

[0108] ;

[0109] where, is the reconstructed node attribute of node p1, is the decoder that reconstructs the node attributes using the potential variable, is the mean of the potential variable in the w-th dimension, , is the value of the potential variable corresponding to node p1 in the w-th dimension, N var is the total dimension of the potential variable, w is the dimension index, is the cubic deviation term;

[0110] Step S35: Prediction; Material requirements have spatio-temporal dependence. The spatio-temporal convolutional layer jointly processes the graph structure and time series, captures cross-material collaboration and time-series lag effects. Mean pooling aggregates multi-node information, suppresses single-material noise, and outputs stable predictions. The second convergence threshold P2 and the maximum number of iterations C2 are preset. Steps S32 - S34 are repeatedly executed until the mean square error between the reconstructed node attributes and the original node attributes is less than or equal to P2 or the maximum number of iterations is reached, at which point the iteration stops. The node attributes U at this time are input into the spatio-temporal convolutional layer to capture spatio-temporal dependence, and then mean pooling is performed to obtain the final node attributes U final , and the production material demand is predicted through a fully connected layer, and the predicted data label is output; the formula used is as follows:

[0111] ;

[0112] where, and are the mean pooling function and the spatio-temporal convolutional layer respectively, P2 = 10 -3 , C2 = 150.

[0113] By performing the above operations, for the problems in the existing production material demand prediction methods that the production material demand is affected by the interaction of many factors, making the material demand exhibit highly non-linear and dynamic changing properties, and traditional methods are difficult to accurately capture these properties, resulting in large demand prediction deviations and low timeliness, this solution combines the node attribute similarity and neighbor difference quantification, calculates the bimodal dynamic weight and updates the node attributes, introduces the second-order perturbation term and the sum of squared deviations, maps the node attributes, corrects the noise distribution dynamically through the gradient, samples the latent variables, introduces the cubic deviation term to reconstruct the node attributes, and obtains the final node attributes through multiple iterations and makes predictions, improving the accuracy of the description of the material relationship, enabling the updated node attributes to better reflect the actual state of the material under complex relationships, mapping the node attributes to the latent space more accurately, the latent variables can better reflect the internal characteristics of the material demand affected by multiple factors, and using the material relationship information more precisely, thereby improving the accuracy and timeliness of the production material demand prediction.

[0114] Example 5, refer to Figure 1 , based on the above example, in step S4, generating a prediction report is to collect and process real-time production material time series data, input it into the production material demand prediction model for prediction, obtain the demand for the corresponding production material according to the output data label, and generate a production material demand prediction report;

[0115] The collection of real-time production material time series data is to collect time stamps, material data, production data, and order data;

[0116] The processing of the real-time production material time series data is to perform data cleaning, data normalization, and data encoding on the real-time production material time series data.

[0117] Example 6, refer to Figure 2 , based on the above example, the production material demand prediction system based on artificial intelligence provided by the present invention includes a production material data integration module, a production material data optimization module, a module for constructing a production material demand prediction model, and a module for generating a prediction report;

[0118] The production material data integration module collects and processes historical production material time series data and sends the data to the production material data optimization module;

[0119] The production material data optimization module clusters the feature vectors and constructs a feature subset, obtains the local classification confidence of the data, combines it with the weight regularization term to construct a loss function, updates the feature vector weights, combines to obtain a normalized feature weight vector, analyzes the feature importance to obtain a smoothing threshold vector, performs feature vector screening based on the two vectors, completes the optimization, and sends the data to the module for constructing a production material demand prediction model;

[0120] The module for constructing a production material demand prediction model combines the node attribute similarity and neighbor difference quantification, calculates the bimodal dynamic weight and updates the node attributes, introduces a second-order perturbation term and the sum of squared deviations, maps the node attributes, dynamically corrects the noise distribution through the gradient, samples the latent variables, introduces a cubic deviation term to reconstruct the node attributes, obtains the final node attributes through multiple iterations and makes a prediction, and sends the data to the module for generating a prediction report;

[0121] The module for generating a prediction report obtains the demand for the corresponding production material according to the data label output by the production material demand prediction model and generates a production material demand prediction report.

[0122] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0123] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention.

[0124] The above description of the present invention and its embodiments is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In general, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, design similar structural forms and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. A method for predicting production material requirements based on artificial intelligence, characterized in that: The method includes the following steps: Step S1: Integrate production material data; collect and process historical production material time series data; Step S2: Optimize production material data; cluster the feature vectors and construct a feature subset to obtain the local classification confidence of the data, combine with the weight regularization term to construct a loss function, update the feature vector weights, combine to obtain a normalized feature weight vector, analyze the feature importance to obtain a smoothing threshold vector, and perform feature vector screening based on the two vectors to complete the optimization; Step S3: Construct a production material demand prediction model; combine the node attribute similarity and neighbor difference quantization, calculate the bimodal dynamic weight and update the node attributes, introduce a second-order perturbation term and the sum of squared deviations, map the node attributes, dynamically correct the noise distribution through the gradient, sample the latent variables, introduce a cubic deviation term to reconstruct the node attributes, and perform multiple iterations to obtain the final node attributes and make predictions; Step S4: Generate a prediction report; obtain the demand for the corresponding production material according to the data label output by the production material demand prediction model, and generate a production material demand prediction report; In step S3, it includes step S34: Reconstruct the node attributes; introduce a cubic deviation term and use the latent variables to reconstruct the node attributes; the formula used is as follows: ； Among them, is the node attribute after the reconstruction of node p1, is the decoder that reconstructs the node attribute using the latent variable, is the mean of the latent variable in the w-th dimension, is the value of the latent variable corresponding to node p1 in the w-th dimension, N var is the total dimension of the latent variable, w is the dimension index, is the cubic bias term, p1 is the node index, x p1 is the latent variable corresponding to node p1, λ is the control factor.

2. The method for predicting production material requirements based on artificial intelligence according to claim 1, characterized in that: In step S3, the construction of the production material demand prediction model specifically includes the following steps: Step S31: Construct an adjacency matrix; use the demand prediction data set as the input data, regard each data in the demand prediction data set as a node, and the node attribute is the corresponding production material time series data. Preset a similarity threshold L. If the cosine similarity between two node attributes is less than the similarity threshold, there is an edge between the corresponding two nodes, and the weight of the edge is 1; otherwise, there is no edge between the corresponding two nodes, and the weight of the edge is 0; obtain the adjacency matrix and perform symmetric normalization processing on the adjacency matrix; Step S32: Update the node attributes; combine the node attribute similarity and neighbor difference quantization, calculate the bimodal dynamic weight between each node and its adjacent nodes, and complete the update of the node attributes based on the bimodal dynamic weight; the formula used is as follows: ； ； Among them, p1, p2, and p3 are node indices, is the node attribute after the update of node p1, s p1 , s p2 and s p3 are the original node attributes of node p1, node p2, and node p3 respectively, is the set of adjacent nodes of node p1, is the bimodal dynamic weight between node p1 and node p2. W and q are learnable weight matrices and weight vectors respectively, and λ is a control factor. and are the node attribute similarity and neighbor difference quantization respectively, is the activation function, is the concatenation operation, T is the transpose operation, and δ is a regulation factor. is the square of the L2 norm; Step S33: Sample the latent variables; Step S34: Reconstruct the node attributes; Step S35: Prediction; preset a second convergence threshold P2 and a maximum number of iterations C2, repeatedly execute Step S32 - Step S34 until the mean square error between the reconstructed node attributes and the original node attributes is less than or equal to P2 or the maximum number of iterations is reached, then stop the iteration. Input the node attributes U at this time into the spatio-temporal convolutional layer to capture spatio-temporal dependencies, and then perform mean pooling to obtain the final node attributes U final , and predict the production material demand through a fully connected layer, and output the predicted data label.

3. The method for predicting production material requirements based on artificial intelligence according to claim 2, characterized in that: In step S33, the sampling of the latent variables; by introducing a second-order perturbation term and the sum of squared deviations, map the node attributes to the mean and variance of the latent space, and dynamically correct the noise distribution through the gradient of the KL divergence loss, and sample the latent variables based on the corrected distribution; the formula used is as follows: ； ； ； where x p1 is the latent variable corresponding to node p1, μ p1 and are the mean and variance respectively that map the attributes of node p1 to the latent space, and are two independent feed-forward neural networks, N att is the total dimension of the node attributes, d is the dimension index, z d is the global attribute mean on the d-th dimension, is the value of the node attribute of node p1 on the d-th dimension, is the small perturbation of the node attribute of node p1 on the d-th dimension, and are the second-order perturbation term and the sum of squared deviations respectively, is the element-wise multiplication operator, is a random noise vector following the standard normal distribution, , N node is the number of nodes, e is the node index, is the standard normal distribution with mean 0 and covariance matrix being the identity matrix I, is the KL divergence loss, is the e-th component of the noise vector, is the partial derivative of with respect to is the partial derivative of network r μ with respect to with respect to.

4. The method for predicting production material requirements based on artificial intelligence according to claim 1, characterized in that: In step S2, the optimization of the production material data specifically includes the following steps: Step S21: Cluster the feature vectors; use the K-means clustering algorithm to divide all the feature vectors in the production material time series data set E into K feature clusters, initialize the weight of each feature vector to 1, and generate an initial feature weight vector of Q dimensions; Step S22: Construct a feature subset; Create A empty sets, for each feature cluster, randomly select feature vectors from it, and sequentially assign them to the A sets, and remove the assigned feature vectors from the feature cluster; If the number of feature vectors in the feature cluster is not zero after removal, evenly and randomly assign the remaining feature vectors in the feature cluster to any one of the sets, and remove the assigned feature vectors from the feature cluster again until the number of feature vectors in each feature cluster is zero, obtaining A feature subsets; where, Z k is the number of feature vectors in the k-th feature cluster, k is the feature cluster index, is the floor operator; Step S23: Update the feature vector weights; Step S24: Feature importance analysis; calculate the Pearson correlation coefficient between every two feature vectors in the production material time series dataset, obtain the redundancy score of each feature vector with the remaining feature vectors, combine the normalized feature weight vector and the redundancy score to generate a deterministic threshold vector, and perform moving average filtering on the deterministic threshold vector to obtain a smoothed threshold vector; Step S25: Feature vector screening; if the value of the v-th column of the normalized feature weight vector h is greater than the value of the v-th column of the smoothed threshold vector, then use the v-th column feature vector corresponding in the production material time series dataset as the effective feature vector; otherwise, use the v-th column feature vector as the redundant feature vector; construct a demand prediction dataset based on all effective feature vectors to complete the optimization.

5. The method for predicting production material requirements based on artificial intelligence according to claim 4, characterized in that: In step S23, the updating of the feature vector weights specifically includes the following steps: Step S231: Calculate the local classification confidence; based on the feature vector weights, calculate the weighted distance between every two data in the production material time series dataset under each feature subset, then introduce a temperature coefficient to control the distribution smoothness, obtain the similarity between every two data, and perform normalization to obtain the local classification confidence of each data under each feature subset; Step S232: Define the loss function; combine the local classification confidence and the weight regularization term to construct the loss function; Step S233: Update; preset a first convergence threshold P1 and a maximum number of updates C1. For each feature subset, use the gradient ascent method to iteratively update the feature vector weights until or stop updating when the maximum number of iterations is reached, and output the feature weight vector at this time; where ω a is the feature weight vector corresponding to the a-th feature subset, is the loss function corresponding to ω a and J max is the maximum loss function value obtained during the current update process. Step S234: Combination; recombine the updated feature weight vectors of A feature subsets according to the index of the initial feature weight vector to obtain a Q-dimensional feature weight vector and perform linear normalization to obtain a Q-dimensional normalized feature weight vector h.

6. The production material demand prediction method based on artificial intelligence according to claim 1, characterized in that: In step S1, the production material data integration is to collect historical production material time series data and process historical production material time series data; The collection of historical production material time series data is to collect timestamps, material data, production data, order data, and demand quantities; The processing of historical production material time series data is to perform data cleaning, data normalization, data encoding, and dataset construction processing on the historical production material time series data to obtain a production material time series dataset.

7. The production material demand prediction method based on artificial intelligence according to claim 1, characterized in that: In step S4, the generation of the prediction report is to collect and process real-time production material time series data, input it into the production material demand prediction model for prediction, obtain the demand quantity of the corresponding production material according to the output data label, and generate a production material demand prediction report; The collection of real-time production material time series data is to collect timestamps, material data, production data, and order data; The processing of real-time production material time series data is to perform data cleaning, data normalization, and data encoding processing on the real-time production material time series data.

8. A production material demand prediction system based on artificial intelligence, used to implement the production material demand prediction method based on artificial intelligence described in any one of claims 1-7, characterized in that: It includes a production material data integration module, a production material data optimization module, a production material demand prediction model construction module, and a prediction report generation module; The production material data integration module collects and processes historical production material time series data and sends the data to the production material data optimization module; The production material data optimization module clusters the feature vectors, constructs feature subsets, obtains the local classification confidence of the data, combines it with the weight regularization term to construct a loss function, updates the feature vector weights, combines them to obtain a normalized feature weight vector, analyzes the feature importance to obtain a smoothing threshold vector, filters the feature vectors based on the two vectors to complete the optimization, and sends the data to the module for constructing the production material demand prediction model; The module for constructing the production material demand prediction model combines the node attribute similarity and neighbor difference quantification, calculates the bimodal dynamic weights and updates the node attributes, introduces the second-order perturbation term and the sum of squared deviations, maps the node attributes, dynamically corrects the noise distribution through the gradient, samples the latent variables, introduces the cubic deviation term to reconstruct the node attributes, obtains the final node attributes through multiple iterations and makes predictions, and sends the data to the module for generating the prediction report; The module for generating the prediction report obtains the demand for the corresponding production material according to the data label output by the production material demand prediction model, and generates a production material demand prediction report.

Images