Data audit analysis method, system and medium for supply chain data elements

Abstract

The present invention relates to a data audit analysis method, system and medium for supply chain data elements. The method includes: performing exploratory data analysis and preprocessing of supply chain data elements; building and training a deep learning model; hybrid integration of models, combining a supply chain demand forecasting model through a hybrid integration strategy and sequential training of integrated pruning; performing online forecasting of multi-step demand values, and performing secure transmission and integrity review of the collected supply chain data. The present invention uses an exploratory data analysis method to mine the autocorrelation, periodicity and trend of supply chain demand in data elements to improve the utilization of supply chain data in the process of model building and demand forecasting. A hybrid integration strategy is designed to further improve the accuracy and robustness of the forecast. Through the supply chain data transmission and review mechanism, with less computing and storage overhead, the security, confidentiality and integrity review of the supply chain data during transmission are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of supply chain data management, and in particular to a data audit and analysis method, system and medium for supply chain data elements. Background Art

[0002] With the rapid development of globalization and information technology, supply chain management has become an indispensable part of modern enterprises. However, the complexity and diversity of supply chain data pose many challenges to data management and analysis. At present, although supply chain management has achieved a certain degree of informatization and automation, there are still many deficiencies in data review and analysis.

[0003] First, supply chain data has the characteristics of large fluctuations and multiple features, which makes it difficult for traditional data analysis methods to effectively deal with it. Each link in the supply chain may generate a large amount of data, including order information, inventory status, transportation trajectory, quality inspection, etc. These data are not only huge in quantity, but also diverse in types, including text, numbers, images and other forms. In addition, due to the dynamic and uncertain nature of the supply chain environment, these data are often highly volatile and unstable, which brings great difficulties to data analysis and mining.

[0004] In terms of data audit and analysis, there is currently a lack of effective methods for further mining and analyzing supply chain data elements based on artificial intelligence algorithms. Although artificial intelligence technology has achieved remarkable application results in other fields, its application in the field of supply chain management is rare. Most of the existing supply chain data analysis methods are based on statistical analysis and simple rule matching, which cannot deeply explore the correlation and regularity behind the data, and it is difficult to meet the needs of modern supply chain management for accurate and efficient data analysis.

[0005] In addition, there is still a lack of research on the security and integrity review mechanism of supply chain data during transmission. In the supply chain, the transmission and sharing of data is inevitable, but this also brings security risks such as data leakage and tampering. Traditional data security protection methods are often difficult to cope with these challenges, and more advanced data encryption, digital signature and other technical means are needed to ensure the security and integrity of data during transmission. Summary of the invention

[0006] In view of the above-mentioned defects in the prior art, the present invention proposes a data audit and analysis method, system and medium for supply chain data elements to achieve data mining, analysis and review for supply chain data elements.

[0007] To achieve the above-mentioned purpose, the present invention adopts the following technical solution: a data review and analysis method for supply chain data elements, comprising:

[0008] Use supply chain data elements to construct a data set as the input data source, and use exploratory data analysis based on the data source to build a feature space, followed by exploratory data analysis and supply chain data element preprocessing;

[0009] Deep learning model construction and training to form LSTM model;

[0010] Model hybrid integration: Through the hybrid integration strategy, several groups of LSTM models are trained offline with different sub-datasets and feature subspaces, and then combined into a supply chain demand forecasting model through sequential training with ensemble pruning;

[0011] The trained supply chain demand forecasting model is used for online forecasting of multi-step demand values, and the collected supply chain data is securely transmitted and integrity reviewed. If it is correctly decrypted and passed the review, a data set is constructed and input into the supply chain demand forecasting model to obtain a forecast of supply chain demand.

[0012] Furthermore, the model hybrid integration includes:

[0013] Furthermore, the exploratory data analysis and supply chain data element preprocessing are specifically:

[0014] Step 1.1: Use exploratory data analysis methods to mine the autocorrelation, periodicity, and trend of supply chain demand in data elements, as well as the correlation between supply chain demand and other external variables. The autocorrelation coefficient is expressed as:

[0015] in is the number of samples in the training data set, is the lag time, For the The sequence value of the moment, For the The sequence value of the moment, represents the sample mean that can be used to approximate the true mean. represents the variance of the sample;

[0016] Step 1.2: According to the correlation factor analysis, a series of external variables are taken as supply chain demand related factors. Then, the data is scaled to the range of [0, 1] using Min-max normalization to make it compatible with the output range of the LSTM activation function and ensure that the contribution of each input data to the output is proportional and standardized. Defined as:

[0017]

[0018] in, Indicates standardized value, and Represents the minimum and maximum values ​​of the time series respectively. At the same time, for a given time period , time-related variables are preprocessed by hot encoding according to hours, weeks, and months.

[0019] Furthermore, the deep learning model construction and training are specifically as follows:

[0020] Step 2.1: Construct an LSTM model, which includes two layers of LSTM networks, with 50 units in the first layer and 100 units in the second layer. A dropout layer is set after each LSTM layer for regularization, and the rates of different dropout layers are set to 0.2 and 0.3 respectively;

[0021] Step 2.2: Design the loss function. The loss function combines the average error of all samples and the maximum error of different sample distributions into coefficients To help model training, and to balance the average error of all samples and the maximum error of different sample distributions, so as to ensure that the model exhibits robust prediction performance in each case, the loss function is:

[0022] in is the actual value, is the predicted value, Auxiliary parameters for model training, represents the average value, Indicates the maximum value;

[0023] Step 2.3: The LSTM model performs forward propagation and backward propagation, and the input layer of the LSTM model obtains a multivariate sequence , is the time step; then, the second layer LSTM outputs a sequence feature vector , represents the time step The hidden state of is the number of units in the second layer; the decoder is set as a fully connected neural network, which transforms the sequence feature vector Pass to Dense layers of output neurons, and generate The predicted value for each time step.

[0024] Furthermore, the model hybrid integration includes:

[0025] Step 3.1: Define the dataset as ,in is the total number of data entries, is the input item, is the corresponding output item; definition is the number of training rounds of the boosting method in LSTM, is the number of training rounds of the random subspace method in LSTM, is the spatial size of the feature subset; first, Rounds of bootstrap sampling are performed to obtain Set training data subset , where the number of samples in each group is the same as the original data set; then, randomly select from the entire feature space of each training set indivual Input variables of size ; create × The training set , to train the base learner , which is expressed as , where the round number subset , Based on the prediction error of the last round of training, the Boosting method is used to improve the accuracy of the model by resampling and adjusting the data distribution. Data subset of round It can be expressed as:

[0026] in , , , is the entire training data set, It is based on Learner exist The mean absolute percentage error on The calculated aggregation coefficient is then converted into the error Exceeding the threshold Sample Append to the training dataset for the next round of bootstrap sampling;

[0027] Step 3.2: After training multiple learners in different feature subspaces in the round, ensemble pruning is performed, and then the optimal learner with the lowest validation error is retained. , for a learner, there exists the following equation:

[0028] in , The validation data set is created by randomly selecting data from the original sample at a fixed ratio of 5%. represents any predicted value of supply chain demand selected from the trained learners; then, the base learners with poor performance are destroyed to obtain a pruned ensemble model with low computational and memory overhead;

[0029] Step 3.3: The output of all base learners is based on the aggregation coefficient For aggregation, The learner selected in the bootstrap sampling round , define its weight for

[0030]

[0031] The model is aggregated according to the above weight coefficients, and a high-performance supply chain demand forecasting model is obtained.

[0032] Furthermore, the secure transmission and integrity review of the collected supply chain data includes:

[0033] Step 4.1: Setup are two prime numbers Factorial cyclic groups, For the group A generator of , defined as For bilinear map operations, four hash functions are defined , , , ,in represents a bit string of arbitrary length, Indicates that the length is fixed to a constant bit string; assuming that all real-time monitoring data will be converted to a length of The bit string , define a filling function , used to convert any length The bit string is filled to Length; extract random number ,in for Integer cyclic groups of order, computing common parameter components , and finally generate public parameters , and the system master key ;

[0034] Step 4.2: Enter system common parameters , supply chain data provider identification code ID, calculate and generate encryption key , enter the system master key and the supply chain data analysis party identification code ID', calculate the decryption key component one , decryption key component 2 、Decryption key component three , and finally generate the decryption key ; For any supply chain data provider, draw a random number As its private key , and calculate the corresponding public key ;

[0035] Step 4.3: Supply chain data provider prepares supply chain data After that, specify the ID code of the supply chain data analyst and extract a random number , calculate the ciphertext component one , ciphertext component 2 , computing ciphertext middleware 1 And ciphertext middleware 2 , based on which the ciphertext component three is calculated ; In addition, calculations on supply chain data Signature , and finally output the supply chain data sign-encrypted ciphertext , and sent to the supply chain data analyst;

[0036] Step 4.4: The supply chain data analyst receives the ciphertext from the supply chain data provider After that, first clarify its identification code ID and calculate the recovery ciphertext middleware And ciphertext middleware 2 ; On this basis, calculate and restore the supply chain data , and then get the supply chain data , using the public key to sign Verify that the following equation is true:

[0037]

[0038] If the equation does not hold, it means that the supply chain data The integrity of the supply chain data is questionable and there is a risk of tampering, so it should be discarded; otherwise, it means that the supply chain data can be used, then on this basis, the data After exploratory data analysis, the supply chain data input set is obtained. ; Input supply chain data into the Input the supply chain demand forecasting model to obtain a forecast of supply chain demand.

[0039] Furthermore, in step 2.1, the activation functions of the LSTM layer in the model are Sigmod and Tanh, while for the dense layer, a linear activation function is used in the last fully connected output layer. .

[0040] Furthermore, in step 2.2, an adaptive learning rate method RMSprop is used as an optimizer in model training.

[0041] Furthermore, at each stage of model training, the loss function is used and RMSprop optimizer to update parameters; the BP algorithm is used in the dense layer and the BPTT algorithm is used in the LSTM layer.

[0042] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the data audit and analysis method for supply chain data elements as described above is implemented.

[0043] The present invention has the following technical effects:

[0044] 1. The present invention uses exploratory data analysis methods to mine the autocorrelation, periodicity, and trend of supply chain demand in data elements, as well as the correlation between supply chain demand and other external variables, so as to improve the utilization of supply chain data in the process of model construction and demand forecasting.

[0045] 2. Considering that the performance of individual networks fluctuates in some cases, the present invention designs a hybrid integration strategy that combines bagging, random subspace and boosting to further improve the accuracy and robustness of prediction.

[0046] 3. The present invention also implements an efficient and secure supply chain data transmission and review mechanism, which uses relatively small computing and storage overhead to achieve the security, confidentiality and integrity review of supply chain data during transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a system architecture diagram of the present invention;

[0048] Figure 2 This is a neural network structure diagram of the present invention;

[0049] Figure 3 It is a main method flow chart of the present invention. DETAILED DESCRIPTION

[0050] The following describes the preferred embodiments of the present invention with reference to the accompanying drawings to make the technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0051] The present invention provides a data audit analysis method for supply chain data elements, which is used to realize data mining, analysis and review for supply chain data elements. The method mainly includes: in the model construction stage, firstly, using the supply chain data elements to construct a data set as the input data source of the model (corresponding to Figure 1 supply chain data provider), and use exploratory data analysis based on raw data to construct feature space (corresponding to Figure 1 Then, the selected data is preprocessed to normalize the data for the next stage of learning model. Then, through the hybrid integration strategy, several groups of LSTM networks are trained offline with different sub-datasets and feature subspaces, and then combined into an integrated model through sequential training of integrated pruning (corresponding to Figure 1 Finally, the trained model is used for online prediction of multi-step demand values. Subsequently, during the model usage phase, the collected supply chain data is securely transmitted and integrity reviewed. If it is correctly decrypted and passed the review, a data set is constructed and input into the model to obtain a forecast of supply chain demand (corresponding to Figure 1 The supply chain demand forecasting model feeds back the demand forecasting results to the supply chain data analysis party).

[0052] The specific steps of the present invention are as follows:

[0053] Step 1: Exploratory data analysis and supply chain data element preprocessing

[0054] Step 1.1: Use exploratory data analysis methods to mine the autocorrelation, periodicity, trend of supply chain demand in data elements, as well as the correlation between supply chain demand and other external variables, to improve the utilization of supply chain data. By analyzing the lagged correlation between current demand and past demand, the number of input time steps can be determined. The autocorrelation coefficient can be expressed as:

[0055] in is the number of samples in the training data set, is the lag time, For the The sequence value of the moment, For the The sequence value of the moment, Represents the sample mean that is a good substitute for the true mean. Represents the sample variance.

[0056] Step 1.2: Supply chain demand has strong randomness and trend characteristics, and is a time series related to many complex factors. In this task, based on the previous analysis of related factors, a series of external variables are used as supply chain demand related factors, such as interest rates and exchange rates. Then, Min-max normalization is used to scale the data to the range of [0,1], making it compatible with the output range of the LSTM activation function and ensuring that each input data contributes proportionally to the output. Standardized It can be defined as:

[0057] in, Indicates standardized value, and Represents the minimum and maximum values ​​of the time series respectively. At the same time, given a time period , time-related variables are preprocessed by hot encoding according to hours, weeks, and months.

[0058] Step 2: Deep learning model construction and training

[0059] Step 2.1: LSTM is a classic and excellent time series network, and is used as a base learner because of its strong ability to process time series data. The LSTM model constructed in the present invention contains two layers of LSTM networks, with 50 units in the first layer and 100 units in the second layer. In order to reduce overfitting, a dropout layer is set after each LSTM layer for regularization, and the rates of different dropout layers are set to 0.2 and 0.3 respectively. The activation functions of the LSTM layer in the model are Sigmod and Tanh. For the dense layer, due to the strong time correlation and tendency of the supply chain demand time series data, as well as the high requirement for prediction accuracy, a linear activation function is used in the last fully connected output layer.

[0060] Step 2.2: During model training, the bias describes the overall accuracy of the model, and the variance describes the stability of the model under different conditions. However, since it is difficult to know the true distribution of supply chain data, the bias and variance cannot be calculated directly. Therefore, inspired by the bias-variance trade-off, the loss function designed in this invention is as follows:

[0061] in is the actual value, is the predicted value. When updating the parameters of the learning algorithm in each batch of training, the loss function should not only pay attention to high accuracy (represented by the first part of the loss function) but also pay attention to strong stability. The loss function combines two different errors (the average error of all samples and the maximum error of different sample distributions) into coefficients To help model training and balance the two errors, the model is guaranteed to show robust prediction performance in every case. For the peak demand prediction scenario, the number of peak demand samples is small, and this loss function can make the prediction error as small as possible. In addition, in model training, the adaptive learning rate method RMSprop is used as an optimizer because it is suitable for handling non-stationary targets and has been proven to be a good recurrent neural network optimizer.

[0062] Step 2.3: The forward propagation and back propagation of the LSTM model in the present invention are as follows Figure 2 As shown. First, the input layer of the model obtains a multivariate sequence , is the time step. Then, the second layer LSTM outputs a sequence feature vector , represents the time step The hidden state of is the number of units in the second layer. The decoder is set as a fully connected neural network, which transforms the sequence feature vector Pass to Dense layers of output neurons, and generate The predicted value of time steps. At each epoch of training, the loss function is used The parameters are updated by using the RMSprop optimizer. The BP algorithm is used in the dense layer and the BPTT algorithm is used in the LSTM layer.

[0063] Step 3: Model hybrid integration

[0064] Step 3.1: Considering that the performance of individual networks fluctuates in some cases, the present invention designs a hybrid integration strategy that combines bagging, random subspace and boosting to further improve the accuracy and robustness of prediction. In the training phase, bagging and random subspace methods increase the diversity of base learners by introducing data randomness through self-bootstrap sampling, and select random feature subspaces from data attributes, which can both enhance diversity and reduce computational costs by reducing the number of input attributes.

[0065] Step 3.2: Define the dataset as ,in is the total number of data entries, is the input item, is the corresponding output item. Definition is the number of training rounds of the boosting method in LSTM, is the number of training rounds of the random subspace method in LSTM, is the spatial size of the feature subset. First, Rounds of bootstrap sampling are performed to obtain Set training data subset , where the number of samples in each group is the same as the original data set. Then, randomly select indivual Therefore, create an input variable of size × The training set , to train the base learner , which is expressed as , where the round number subset , Based on the prediction error of the last round of training, the Boosting method is used to improve the accuracy of the model by resampling and adjusting the data distribution. Data subset of round It can be expressed as:

[0066] in , , , is the entire training data set, It is based on Learner exist The mean absolute percentage error on The calculated aggregation coefficient. Exceeding the threshold Sample Append to the training dataset for the next round of bootstrap sampling.

[0067] Step 3.3: After training multiple learners in different feature subspaces in a round, ensemble pruning is performed, and then the best learner with the lowest validation error is retained. , for a learner, there exists the following equation:

[0068]

[0069] in , A validation dataset is created by randomly selecting data from the original sample at a fixed ratio of 5%. represents any predicted value of supply chain demand selected from the trained learners. Then, the base learners with poor performance are destroyed to obtain a pruned ensemble model with low computational and memory overhead.

[0070] Step 3.4: The output of all base learners is based on the aggregation coefficient For aggregation. The learner selected in the bootstrap sampling round , define its weight for

[0071]

[0072] The model is aggregated according to the above weight coefficients, and a high-performance supply chain demand forecasting model is obtained. The following steps introduce the safe use of the model.

[0073] Step 4: Supply Chain Data Security Transmission and Integrity Review

[0074] This step is used to conduct an integrity review of the collected supply chain data to ensure data quality and prediction results. On this basis, the above-trained model and data are used to predict the demand of the supply chain. Specifically, it is divided into the following steps:

[0075] Step 4.1: Setup are two prime numbers Factorial cyclic groups, For the group A generator of , defined as For bilinear mapping operations. Define four hash functions , , , ,in represents a bit string of arbitrary length, Indicates that the length is fixed to a constant Assume that all real-time monitoring data will be converted to a length of The bit string , define a filling function , used to convert any length The bit string is filled to Length. Extract random number ,in for Integer cyclic groups of order, computing common parameter components . Finally generate public parameters , and the system master key .

[0076] Step 4.2: Enter system common parameters , supply chain data provider identification code ID, calculate and generate encryption key . Enter the system master key and the supply chain data analysis party identification code ID', calculate the decryption key component one , decryption key component 2 、Decryption key component three , and finally generate the decryption key For any supply chain data provider, draw a random number As its private key , and calculate the corresponding public key .

[0077] Step 4.3: Supply chain data provider prepares supply chain data After that, specify the ID code of the supply chain data analyst and extract a random number , calculate the ciphertext component one , ciphertext component 2 , computing ciphertext middleware 1 And ciphertext middleware 2 On this basis, the ciphertext component three is calculated In addition, calculations on supply chain data Signature The final output is the supply chain data signcryption ciphertext , and sent to the supply chain data analyst.

[0078] Step 4.4: The supply chain data analyst receives the ciphertext from the supply chain data provider After that, first clarify its identification code ID and calculate the recovery ciphertext middleware And ciphertext middleware 2 On this basis, the supply chain data is calculated and restored , and then get the supply chain data . Use the public key to sign Verify that the following equation is true:

[0079]

[0080] If the equation does not hold, it means that the supply chain data The integrity of the data is questionable and there is a risk of tampering, so it should be discarded. Otherwise, it means that the supply chain data can be used, then on this basis, the data After the exploratory data analysis method in step 1, the supply chain data input set is obtained. . Input supply chain data into the Input the supply chain demand forecasting model obtained in step three to obtain a forecast of supply chain demand.

[0081] In view of the problems of the prior art, the key technical problems to be solved by the present invention are: (1) How to establish a supply chain demand forecasting model based on supply chain data elements; (2) Supply chain data usually fluctuates and changes greatly, how to improve the performance and accuracy of the model in such cases; (3) Supply chain data is vulnerable to security threats during transmission, how to achieve secure transmission and review. Through the above method, the following technical effects are achieved:

[0082] 1. The present invention uses exploratory data analysis methods to mine the autocorrelation, periodicity, and trend of supply chain demand in data elements, as well as the correlation between supply chain demand and other external variables, so as to improve the utilization of supply chain data in the process of model construction and demand forecasting.

[0083] 2. Considering that the performance of individual networks fluctuates in some cases, the present invention designs a hybrid integration strategy that combines bagging, random subspace and boosting to further improve the accuracy and robustness of prediction.

[0084] 3. The present invention also implements an efficient and secure supply chain data transmission and review mechanism, which uses relatively small computing and storage overhead to achieve the security, confidentiality and integrity review of supply chain data during transmission.

[0085] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A data audit analysis method for supply chain data elements, characterized in that: include: Use supply chain data elements to construct a data set as the input data source, and use exploratory data analysis based on the data source to build a feature space, followed by exploratory data analysis and supply chain data element preprocessing; Deep learning model construction and training to form LSTM model; Model hybrid integration: Through the hybrid integration strategy, several groups of LSTM models are trained offline with different sub-datasets and feature subspaces, and then combined into a supply chain demand forecasting model through sequential training with ensemble pruning; The trained supply chain demand forecasting model is used for online forecasting of multi-step demand values. The collected supply chain data is securely transmitted and integrity reviewed. If it is correctly decrypted and reviewed, a data set is constructed and input into the supply chain demand forecasting model to obtain a forecast of supply chain demand. The exploratory data analysis and supply chain data element preprocessing include: Step 1.1: Use exploratory data analysis methods to mine the autocorrelation, periodicity, and trend of supply chain demand in data elements, as well as the correlation between supply chain demand and other external variables. The autocorrelation coefficient is expressed as: Where N is the number of samples in the training data set, l is the lag time, and x i is the sequence value at the i-th moment, x i+l is the sequence value at the i+lth moment, represents the sample mean that can be used to approximate the true mean. represents the variance of the sample; Step 1.2: According to the correlation factor analysis, a series of external variables are taken as supply chain demand related factors. Then, the data is scaled to the range of [0, 1] using Min-max normalization to make it compatible with the output range of the LSTM activation function and ensure that the contribution of each input data to the output is proportional. The standardized x i Defined as: in, represents the standardized x i Value, x min and x max They represent the minimum and maximum values ​​of the time series respectively. At the same time, given a time period i, the time-related variables are preprocessed by hot encoding according to hours, weeks, and months; The deep learning model construction and training are specifically as follows: Step 2.1: Construct an LSTM model, which includes two layers of LSTM networks, with 50 units in the first layer and 100 units in the second layer. A dropout layer is set after each LSTM layer for regularization, and the rates of different dropout layers are set to 0.2 and 0.3 respectively; Step 2.2: Design a loss function. The loss function helps model training by combining the average error of all samples and the maximum error of different sample distributions into a coefficient e, and makes a trade-off between the average error of all samples and the maximum error of different sample distributions, thereby ensuring that the model exhibits robust prediction performance in each case. The loss function is: Where y is the actual value, is the predicted value, e is the auxiliary parameter of model training, mean 0<i≤N Indicates the average value, max 0<i≤N Indicates the maximum value; Step 2.3: The LSTM model performs forward propagation and backward propagation, and the input layer of the LSTM model obtains a multivariate sequence {X T-m , X T-m+1 ,...,X T }, m is the time step; then, the second layer LSTM outputs a sequence feature vector H T ={F1, F2, ..., F U }, represents the hidden state of the time step T, where U is the number of units in the second layer; the decoder is set as a fully connected neural network, which transforms the sequence feature vector H T Passed to a dense layer with n output neurons and generates predictions for n time steps; The model is hybrid integrated, including: Step 3.1: Define the dataset as D={(x1, y1), (x2, y2),..., (x i ,y i ),...,(x m ,y m )}, where m is the total number of data items, x is the input item, and y is the corresponding output item; define T1 as the number of training rounds of the boosting method in LSTM, T2 as the number of training rounds of the random subspace method in LSTM, and d as the spatial size of the feature subset; first, bootstrap sampling is performed on the T1 round to obtain the T1 group training data subset The number of samples in each group is the same as that in the original data set. Then, d T2 input variables are randomly selected from the entire feature space of each training set to create a training set of size T1×T2. To train the base learner That is, it is expressed as The round number subsets are t1∈T1, t2∈T2; based on the prediction error of the last round of training, the Boosting method is used to improve the accuracy of the model by resampling and adjusting the data distribution. The data subset of the t1th round It can be expressed as: in D is the entire training data set, α j is based on the jth learner The mean absolute percentage error on D The calculated aggregation coefficient is then converted into the error Samples (x i ,y i ) is appended to the training dataset for the next round of bootstrap sampling; Step 3.2: After training multiple learners in different feature subspaces in round t1, perform ensemble pruning and retain the best learner with the lowest validation error. Given t1∈T1, for a learner, there exists the following equation: , where t2∈T2, V is a validation dataset created by randomly selecting data from the original sample at a fixed ratio of 5%, represents any predicted value of supply chain demand selected from the trained learners; then, the base learners with poor performance are destroyed to obtain a pruned ensemble model with low computational and memory overhead; Step 3.3: The output of all base learners is based on the aggregation coefficient α j Aggregate, for the learner selected in the bootstrap sampling of round t1 Define its weight for The model is aggregated according to the above weight coefficients, and a high-performance supply chain demand forecasting model is obtained; The secure transmission and integrity review of the collected supply chain data includes: Step 4.1: Set G, G T are two prime numbers q-order multiplicative cyclic groups, P is a generator of group G, and we define e: G×G→G T For bilinear mapping operations, four hash functions H1 are defined: {0, 1} * →G, H2: {0, 1} * →G, H3:G T →{0, 1} i , H4: {0, 1} * →G, where {0, 1} * Represents a bit string of arbitrary length, {0, 1} l represents a bit string with a fixed length of constant l; assuming that all real-time monitoring data m will be converted into a bit string {0, 1} with a length of n n , define a filling function Φ: {0, 1} n →{0, 1} l , used to fill any bit string of length n to length l; extract random numbers in For a q-order integer cyclic group, calculate the common parameter component P0 = P r , and finally generate the public parameter mpk = (e, G, G T , q, P, P0, H1, H2, H3, H4, Φ), and system master key msk=(r, s); Step 4.2: Input the system public parameter mpk, the supply chain data provider identification code ID, and calculate the encryption key ek ID =H2(ID) s , input the system master key msk and the supply chain data analysis party identity code ID′, calculate the decryption key component one Decryption key component 2 Decryption key component three Finally generate the decryption key For any supply chain data provider, a random number r is drawn id ∈G as its private key sk ID , and calculate the corresponding public key Step 4.3: Supply chain data provider prepares supply chain data m∈{0,1} n After that, specify the ID' of the supply chain data analyst and extract a random number Calculate the ciphertext component T = P t , ciphertext component 2 U=P u , computing ciphertext middleware 1 And the ciphertext middleware 2k S =e(H1(ID′),T·ek ID ), and on this basis, calculate the ciphertext component three In addition, the signature of the supply chain data m is calculated Finally, the supply chain data signcryption ciphertext C = (T, U, V, σ) is output and sent to the supply chain data analysis party; Step 4.4: After receiving the ciphertext C = (T, U, V, σ) from the supply chain data provider, the supply chain data analyst first determines its identification code ID and calculates the ciphertext recovery middleware And ciphertext middleware 2 On this basis, the supply chain data is calculated and restored Then, the supply chain data m is obtained, and the signature σ is verified using the public key, that is, whether the following equation is true: e(σ, g) = e(H4(m), pk ID ) If the equation does not hold, it means that the integrity of the supply chain data m is in doubt and there is a risk of being tampered with, so it should be discarded; otherwise, it means that the supply chain data m can be used. On this basis, the data m is processed by an exploratory data analysis method to obtain a supply chain data input set D′; the supply chain data input set D′ is input into the supply chain demand forecasting model to obtain a forecast of the supply chain demand; In step 2.1, the activation functions of the LSTM layer in the model are Sigmod and Tanh, and for the dense layer, a linear activation function f(x)=x is used in the last fully connected output layer.

2. According to the data audit and analysis method for supply chain data elements according to claim 1, in step 2.2, the adaptive learning rate method RMSprop is used as the optimizer in model training.

3. According to the data audit analysis method for supply chain data elements of claim 1, in each period of model training, a loss function is used and RMSprop optimizer to update the parameters; among them, The BP algorithm is used in the dense layer, and the BPTT algorithm is used in the LSTM layer.

4. A computer system, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the data audit and analysis method for supply chain data elements as described in any one of claims 1 to 3.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a data audit and analysis method program for supply chain data elements, and when the data audit and analysis method program for supply chain data elements is executed by a processor, the steps of the method described in any one of claims 1 to 3 are executed.

Images