An auto parts fragile part store goods purchase recommendation method in a big data scenario

Abstract

The application provides a method for recommending purchases of auto parts and vulnerable parts stores in a big data scenario, comprising the following steps: S10: preprocessing historical purchase, sales and inventory data of the stores to build a training data set; S20: building a basic prediction model according to the training data set; S30: building an integrated prediction model according to the basic prediction model to determine a sales prediction value of the stores; S40: building a category safety inventory water level model according to sales and inventory data of stores nationwide in combination with the sales prediction value of the stores; S50: building a product package clustering model according to historical sales data of single-store goods to divide the goods into product packages; and S60: determining recommended purchase values of store goods and product package attributions according to the integrated prediction model, the safety inventory water level model and the clustering analysis model, so that the method is more scientific in purchase, reduces inventory accumulation under the condition of sufficient goods and is beneficial to inventory turnover.

Description

TECHNICAL FIELD

[0001] The application relates to a method for recommending purchases of auto parts and consumables by stores, and applies to the field of purchase of auto parts and consumables by stores. BACKGROUND

[0002] In the daily operation of stores of auto parts and consumables, the steps of preparing goods and purchasing goods have an important influence on the sales, inventory, and capital turnover of the stores. Store purchasers integrate and process the quotation documents of all automobile parts suppliers and make comprehensive considerations in multiple aspects. After considering factors such as inventory, price, delivery efficiency, and part quality, the store purchaser initiates a purchase request to the automobile parts supplier that best meets the requirements. After receiving the purchase request from the store purchaser, the automobile parts supplier opens a sales order and arranges delivery. The traditional automobile parts purchase transaction mode has a complex operation process and consumes a lot of time. In the field of auto parts, the sku data is large, and the store purchaser needs to have a good understanding of the market and the store operation to ensure the rationality of the purchases. Since auto parts and consumables are products needed to solve safety problems for automobiles, if the inventory is insufficient and the customer cannot immediately replace the parts, the customer will be lost. If the inventory is accumulated, the goods and capital turnover will be overloaded, which may lead to poor operation. The traditional automobile parts and consumable purchase mode consumes a large amount of enterprise resources and human cost, resulting in a complex and inefficient operation process and high cost.

[0003] Therefore, the research purpose of the present application is to design a method for recommending purchases of auto parts and consumables by stores in a big data scenario to solve the above problems. SUMMARY

[0004] The present application provides a method for recommending purchases of auto parts and consumables by stores in a big data scenario, which can effectively solve the above problems.

[0005] The present application is implemented as follows:

[0006] A method for recommending purchases of auto parts and consumables by stores in a big data scenario, comprising the following steps:

[0007] S10: Preprocessing the historical purchase, sale, and inventory data of stores of auto parts and consumables to construct a training data set;

[0008] S20: Constructing a basic prediction model according to the training data set;

[0009] S30: Constructing an integrated prediction model according to the basic prediction model to determine the sales prediction value of the store, comprising:

[0010] According to the training data set, a first basic prediction model is constructed by using linear regression to fit the linear relationship between the feature set and the sales.

[0011] According to the training data set, a second basic prediction model is constructed using random forest regression to explore the nonlinear relationship between the feature set and the sales;

[0012] According to the training data set, a third basic prediction model is constructed using gradient boosting regression;

[0013] S40: According to the sales and inventory data of the national auto parts consumable parts stores, combined with the sales prediction value of the auto parts consumable parts stores, a category safety inventory level model is constructed, including:

[0014] The sales and inventory data of the stores in the past three months are summarized and calculated, and the top 20% of stores in terms of sales quantity and stock-to-sales ratio in the category are selected. This set represents stores with excellent sales operation and inventory management capabilities in each category;

[0015] The overall category inventory month of the store set is calculated as a reference for the category turnover stock of the national stores;

[0016] S50: According to the historical sales data of individual auto parts consumable parts stores, a product package clustering model is constructed to divide products into product packages, including:

[0017] A KMeans algorithm is used to construct a product package clustering model to provide a purchase priority reference for service providers. The input data set of the model selects the sales quantity, gross profit amount, and sales frequency features of the store single product as features. The algorithm training process is as follows:

[0018] S501: Randomly select k data points as the center points of k clusters;

[0019] S502: For each data point, calculate the Euclidean distance between the data point and the k center points, and divide the data point into the cluster represented by the shortest center point;

[0020] S503: After all data points are divided, the mean of the data points in the cluster is used to update the center points of the k clusters;

[0021] S504: Repeat S502 and S503 above until the center points no longer change, and the training stops;

[0022] S60: According to the integrated prediction model, the safety inventory level model, and the clustering analysis model, the recommended purchase value of the auto parts consumable parts store goods and the product package attribution are calculated;

[0023] According to the integrated prediction model, the sales prediction value of the store single product is determined, the recommended purchase value is output through the category safety inventory level model combined with the current inventory of the store, and the product purchase priority is output combined with the product package clustering model.

[0024] As a further improvement, the training data set contains data of multiple commodities, for any commodity, a data set D = {( , ), …, ( , )} corresponding to it; is the feature data set of the nth sample, is the next month's sales volume of the nth sample, the feature data includes the sales volume, the purchase quantity, the number of customers, the average purchase price of the month, the number of newly developed customers, the single price of the repair factory, the purchase of active service providers, and the purchase of non-active first batch service providers in the past preset number of months.

[0025] As a further improvement, The first basic prediction model is constructed by using linear regression to fit the linear relationship between the feature set and the sales volume according to the training data set, which includes:

[0026] The constructed linear regression model is:

[0027]

[0028] Then, the parameter vector θ is obtained by solving the minimum value of the loss function:

[0029] The gradient descent method is used to solve the linear regression:

[0030]

[0031] Where x is the input sample feature set, θ is the model parameter, h(x) is the model prediction value, y is the sample actual value, a is the learning rate, and J(θ) is the loss function.

[0032] As a further improvement, the second basic prediction model is constructed by using random forest regression to explore the non-linear relationship between the feature set and the sales volume according to the training data set, which includes:

[0033] Using resampling, m samples are randomly selected from the training data set to obtain a new sub-data set;

[0034] Three features are randomly selected in the sub-data set to train a CART regression tree;

[0035] The above two steps are repeated n times to construct a random forest model composed of n regression trees; the prediction value of the random forest is determined by the prediction results of all regression trees;

[0036] The generation of the CART regression tree is the process of recursively constructing a binary tree, and each step of generating the tree is to solve the following equation:

[0037]

[0038] Iterate through any feature A and any dividing point s in the dataset to divide the dataset into subsets. and Requesting an envoy and Each set has the smallest mean squared error, and at the same time and The feature A and the split point s corresponding to the minimum sum of the mean squared errors are used as branches to generate the current tree.

[0039] As a further improvement, the construction of the third base prediction model using gradient boosting regression based on the training dataset includes:

[0040] S201: The gradient boosting regression model uses a regression tree as a base learner and employs the Boosting approach, using the negative gradient of the loss function as a residual fitting method to obtain multiple weak learners. Finally, the strong learners are integrated to construct the gradient boosting regression model.

[0041] Where x is the input sample, w is the model parameter, h is the regression tree, a is the weight of each tree, and m represents the m-th regression tree;

[0042] S202: Constructing the initial weak learner:

[0043] S203: Solve for the loss function, set the derivative to 0, and obtain the initial weak learner predictions: = AVG( );

[0044] Where y is the actual value of the sample, c is the predicted value of the model, L(y,c) is the variance and loss function, and i represents the i-th sample data;

[0045] S204: Calculate the negative gradient, i.e., the residual: - The fitted object for the next weak learner is used to construct a regression tree. All features and split points are traversed to find the split point with the smallest residual SSE after splitting, and the tree is generated.

[0046] S205: Repeat S203 above to obtain the prediction value of the weak learner, and update the strong learner using the learning rate; = + a* , where a is the learning rate, and S204 is repeated to obtain the residuals for the next model;

[0047] S206: Repeat S205 until the required number of iterations is met, and finally solve for the strong learner.

[0048] As a further improvement, the integrated prediction model is constructed according to the basic prediction model, and the determination of the sales prediction value of the store comprises:

[0049] The integrated prediction model is constructed according to the first basic prediction model, the second basic prediction model and the third basic prediction model: W(x) = a1F1(x) + a2F2(x) + a3F3(x) (x) + (x) + (x);

[0050] Wherein, F(x) is a basic prediction model, a is the weight of the basic prediction model, and W(x) is an integrated prediction model;

[0051] For each product data set of the store, the training data set is divided, and the first three months of the prediction month are taken as the boundary to divide into a sub-training set and a sub-test set. By evaluating the error rate of the sub-test set, the weights of the K basic prediction models are determined ;

[0052] Solving: = / ∑ , is the error rate of the i-th basic prediction model on the sub-test set.

[0053] The beneficial effects of the present application are:

[0054] (1) According to the training data set, a basic prediction model is constructed, and the basic prediction model is constructed into an integrated prediction model to determine the sales prediction value of the store. For each product data set of the store, the training data set is divided, and the first three months of the prediction month are taken as the boundary to divide into a sub-training set and a sub-test set. By evaluating the error rate of the sub-test set, the weights of the three basic prediction models are determined, and the integrated prediction model has a lower error rate than the basic prediction model, making the prediction more accurate.

[0055] (2) According to the sales and inventory data of the stores nationwide, combined with the sales prediction value of the store, a category safety stock level model is constructed. The sales and inventory data of the stores in the past three months are aggregated and calculated, and the top 20% of the stores in the sales quantity and the sales ratio of the category are selected to calculate the overall category inventory month as a reference for the category turnover of the stores nationwide. According to the market and inventory sales data, the category safety stock level model is updated in real time, and the recommended purchase value is output. By screening the sales quantity in the category, the recommended purchase value is more accurate.

[0056] (3) According to the single store commodity sales forecast value, a commodity bag clustering model is constructed to divide the commodities into commodity bags; the KMeans algorithm is used to construct the commodity bag clustering model to provide the service provider with a purchase priority reference, the feature of the input data set of the model is selected as the sales quantity, gross profit amount, and sales frequency feature, the commodity bag clustering model is trained and constructed to provide the service provider with a purchase priority reference, accurately predict the store purchase priority, adapt to the characteristics of each store, provide a more suitable store purchase plan to make the purchase more scientific, and reduce the accumulation of inventory under the condition of sufficient goods, which is beneficial to inventory turnover. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation to the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0058] Figure 1 is a step flow chart provided by an embodiment of the present application.

[0059] Figure 2 is a step program block diagram provided by an embodiment of the present application.

[0060] Figure 3 is a whole average error line chart of the prediction results of all commodities of the store provided by an embodiment of the present application.

[0061] Figure 4 is a whole average error pie chart of the prediction results of all commodities of the store provided by an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0063] In the description of the present application, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise explicitly specified.

[0064] Referring to Figures 1-2 As shown in the figure, a method for recommending purchases of auto parts and vulnerable parts stores in a big data scenario includes the following steps:

[0065] S10: Preprocess the historical purchase, sales and inventory data of the store to build a training data set; the training data set contains data for multiple goods, and for any good, a data set D = {( , ), …, ( , )} corresponding to the data set; is the feature data set of the i-th sample, is the sales volume of the i-th sample next month, and the feature data includes the sales volume, purchase volume, number of sales customers, average purchase price this month, number of new customers, single price of repair shop, purchase of active service providers, and purchase of non-active non-first batch service providers in the past preset number of months.

[0066] The main process of data preprocessing for the data set includes:

[0067] (1) Complete the missing months of the data. The advantage of completing the missing months of the data is to make the data more complete, and to reduce the error rate caused by incomplete data during data processing.

[0068] (2) Use feature engineering to build derived features of sales volume based on the autoregressive idea, use these additional features to improve the quality of the results of the machine learning process, use the performance of the same variable in previous periods to predict the performance of the variable in the current period, and assume that they are in a linear relationship.

[0069] (3) Use the Isolation Forest algorithm to identify and exclude abnormal data, exclude abnormal data, and reduce the situation of large data errors caused by abnormal data.

[0070] S20: Building a basic prediction model according to the training data set; the building of the basic prediction model according to the training data set includes:

[0071] Using linear regression to fit the linear relationship between the feature set and the sales volume according to the training data set to build a first basic prediction model includes:

[0072] The constructed linear regression model is as follows:

[0073]

[0074] Then, by finding the minimum value of the loss function, the parameter vector θ is obtained;

[0075] Solving linear regression using gradient descent:

[0076]

[0077] Where x is the set of input sample features, θ is the model parameters, h(x) is the model prediction value, y is the actual sample value, a is the learning rate, and J(θ) is the loss function.

[0078] Based on the training dataset, a second basic prediction model is constructed by using random forest regression to explore the non-linear relationship between the feature set and sales volume, including:

[0079] A new subset of the dataset is obtained by randomly selecting m samples from the training dataset using a resampling method.

[0080] Three features are randomly selected from a subset of the dataset, and a CART regression tree is trained and generated.

[0081] Repeat the above two steps n times to construct a random forest model consisting of n regression trees; the predicted value of the random forest is determined by the prediction results of all regression trees.

[0082] The generation of a CART regression tree is a recursive process of constructing a binary tree. Each step in generating the tree involves solving the following equation:

[0083]

[0084] Iterate through any feature A and any dividing point s in the dataset to divide the dataset into subsets. and Requesting an envoy and Each set has the smallest mean squared error, and at the same time and The feature A and the split point s corresponding to the minimum sum of the mean squared errors are used as branches to generate the current tree.

[0085] Based on the training dataset, constructing a third basic prediction model using gradient boosting regression includes:

[0086] S201: The gradient boosting regression model uses regression trees as base learners, adopts the boosting idea, uses the negative gradient of the loss function as the residual fitting method, obtains multiple weak learners, and finally integrates a strong learner. The gradient boosting regression model is constructed:

[0087] wherein x is an input sample, w is a model parameter, h is a regression tree, a is the weight of each tree, and m represents the mth regression tree;

[0088] S202: Construct an initial weak learner:

[0089] S203: Solve the minimum loss function, set the derivative equal to 0, and obtain the initial weak learner prediction value: = AVG( );

[0090] wherein y is the actual value of the sample, c is the model prediction value, L(y, c) is the variance and loss function, and i represents the ith sample data;

[0091] S204: Calculate the negative gradient, i.e. the residual: - , which is used to construct a regression tree for fitting the next weak learner, traverse all features and split points, find the split point with the minimum residual SSE after splitting, and generate the tree;

[0092] S205: Repeat the above S203 to obtain the weak learner prediction value, and update the strong learner using the learning rate; = + a* , a is the learning rate, and the residual for the next model is obtained by repeating S204;

[0093] S206: Repeat S205 until the number of iterations is satisfied, and finally solve the strong learner.

[0094] S30: Construct an integrated prediction model according to the basic prediction model, and determine the sales prediction value of the store, comprising:

[0095] According to the first basic prediction model, the second basic prediction model, and the third basic prediction model, an integrated prediction model is constructed: W(x) = (x) + (x) + (x);

[0096] wherein F(x) is a basic prediction model, a is the weight of the basic prediction model, and W(x) is an integrated prediction model;

[0097] For each product data set of the store, the training data set is divided to predict the first three months of the month as the boundary, and divided into sub-training set and sub-test set. By evaluating the error rate of the sub-test set, the weight of the three basic prediction models is determined ; solve: = / ∑ , is the error rate of the i-th basic prediction model on the sub-test set. In the daily replenishment and purchase link of the auto parts consumable store, due to the large amount of preprocessed feature data, including past sales, purchase quantity, number of sales customers, monthly purchase average, new customer number, repair shop single price, active service provider purchase, non-active non-first batch service provider purchase, in order to ensure the accuracy of the replenishment and purchase, it is necessary to reduce the data error rate when processing data, to avoid unscientific replenishment and purchase caused by large data error, as shown in Figures 3-4 , the error rate of the integrated prediction model is lower than that of the basic prediction model, and the prediction is more accurate.

[0098] S40: According to the sales and inventory data of the national stores, combined with the sales prediction value of the store, a category safety stock level model is constructed, including:

[0099] The sales and inventory data of the stores in the past three months are summarized and calculated, and the top 20% of the stores in the sales quantity and the sales ratio of the category are selected. The set represents the stores with excellent sales operation and inventory management capability in each category. For the store set, the overall category inventory month is calculated as the reference for the category turnover of the national stores. Since customers value the cost performance or appearance of the auto parts consumables, according to the sales data of the stores with excellent sales operation and inventory management capability in each category, the market-oriented excellent auto parts consumables and auto parts manufacturers are reflected. Selecting the top 20% of the stores in the sales quantity and the sales ratio of the category is because the top 20% of the stores can reflect the market orientation of the auto parts market. According to the market and inventory sales data, the category safety stock level model is updated in real time, and the recommended purchase value is output. By screening the sales quantity in the category, the recommended purchase value is more accurate.

[0100] S50: According to the historical sales data of single-store goods, a goods package clustering model is constructed to divide the goods into goods packages, including:

[0101] A KMeans algorithm is used to construct a goods package clustering model to provide a purchase priority reference for service providers. The feature set of the input data set of the model selects the sales quantity, gross profit amount, and sales frequency characteristics of the single product of the store. The algorithm training process is as follows:

[0102] S501: initialize randomly select K data points as the center of K clusters, in this embodiment, the K = 3, that is, select 3 data points as the center of 3 clusters;

[0103] S502: for each data point, calculate the Euclidean distance between the data point and the 3 center points, divide the data point to the cluster represented by the shortest center point; by calculating the distance between each object and each seed cluster center, each object is assigned to the nearest cluster center, and the cluster center and the object assigned to it represents a cluster.

[0104] S503: after all data points are divided, update the 3 cluster centers using the mean of the data points within the cluster;

[0105] S504: repeat the above S502 and S503 until the center point no longer changes, the training stops, the cluster center of the cluster is recalculated according to the existing objects in the cluster every time a sample is assigned, and this process is repeated until a certain termination condition is met, the termination condition can be that no object is reassigned to a different cluster, no cluster center changes, and the error sum of squares is locally minimized.

[0106] The commodity package clustering model is trained, and the commodities are divided into 3 clusters, the cluster with the highest sales amount is named as the mandatory package, and the remaining two clusters are named as the recommended package, in this embodiment, the remaining two clusters named as the recommended package are integrated to provide the priority of purchasing goods for the service provider.

[0107] As shown in Table 1, the clustering model result of the sample store is: the priority of purchasing goods is divided into two levels, one is the mandatory package, and the mandatory package is the highest item in the average sales gross profit, average sales days, and average sales quantity, which must be purchased when preparing to purchase goods; one is the recommended package, and the system recommends selective purchase, the KMeans algorithm is used to construct the commodity package clustering model to provide the priority of purchasing goods, which is suitable for the characteristics of different stores, provides a more suitable store purchasing scheme to make the purchasing more scientific, reduces the accumulation of inventory in the case of sufficient goods, and is beneficial to inventory turnover.

[0108] Table 1 Clustering model result of sample store

[0109]

[0110] S60: According to the integrated prediction model, the safety stock water level model and the clustering analysis model, the recommended purchase value of the store goods and the commodity package attribution are determined, including: according to the integrated prediction model, the sales prediction value of the store single product is determined, the recommended purchase value is output through the category safety stock water level model combined with the current inventory of the store, and the commodity purchase priority is output combined with the commodity package clustering model; there are countless components on the car, some of which have a short service life, and if they are not replaced in time, not only will the vehicle be damaged, but also the driving safety of the owner and passengers will be threatened, therefore, people pay more and more attention to the maintenance of the car and the replacement of the automobile parts, since the SKU quantity of the auto parts is large, in the daily preparation and purchase process, the general store operator adopts subjective feeling or checks simple sales data to judge whether the current inventory is sufficient, and decides the commodity package and quantity of purchase, which leads to unscientific purchase behavior, and further affects the store gross profit and inventory turnover, according to the integrated prediction model in the application, the error rate is reduced compared with the basic prediction model, the prediction accuracy is effectively improved, the category safety stock water level model is constructed according to the sales and inventory data of the national stores combined with the sales prediction value of the store; a store set with sales quantity and sales ratio in the top 20% of the category is selected, the overall category inventory month is calculated, and the category turnover preparation of the national stores is used as a reference; the category safety stock water level model is updated in real time according to the market and inventory sales data, the recommended purchase value is output, and the recommended purchase value is more accurate by screening the sales quantity in the category; the commodity package clustering model is constructed by using the KMeans algorithm, the purchase priority reference is provided for the service provider, the feature of the input data set of the model is selected as the sales quantity, gross profit amount and sales frequency feature, the commodity package clustering model is trained and constructed, the purchase priority reference is provided for the service provider, the store purchase priority is accurately predicted, the characteristics of various stores are suitable, a more suitable store purchase scheme is provided, so that the purchase is more scientific, in the case of sufficient goods, the inventory accumulation is reduced, which is beneficial to the inventory turnover.

[0111] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for recommending the purchase of easily damaged auto parts in a big data scenario, characterized in that, The method comprises the following steps: S10: preprocessing historical sales and inventory data of auto parts stores, and constructing a training data set; S20: constructing a basic prediction model according to the training data set, comprising: using linear regression to fit the linear relationship between the feature set and the sales according to the training data set, and constructing a first basic prediction model; using random forest regression to explore the nonlinear relationship between the feature set and the sales according to the training data set, and constructing a second basic prediction model; constructing a third basic prediction model using gradient boosting regression according to the training data set; constructing an integrated prediction model according to the first, second and third basic prediction models; S30: constructing an integrated prediction model according to the basic prediction model to determine the sales prediction value of the store; S40: constructing a category safety stock level model according to the sales and inventory data of auto parts stores nationwide, and combining the sales prediction value of the auto parts stores, comprising: performing category aggregation and calculation on the sales and inventory data of stores in the past three months, and selecting a store set with the top 20% of sales quantity and stock-to-sales ratio in the category, which represents stores with excellent sales operation and inventory management capabilities in each category; calculating the overall category inventory month of the store set as a reference for category turnover stock of stores nationwide; S50: constructing a product package clustering model according to the historical sales data of individual auto parts stores, and dividing products into product packages, comprising: using the KMeans algorithm to construct a product package clustering model to provide a priority reference for the service provider, and the input data set of the model is selected from the sales quantity, gross profit amount, and sales frequency features of the store's single product, and the algorithm training process is as follows: S501: randomly select p data points as the center points of p clusters; S502: for each data point, calculate the Euclidean distance between the data point and the p center points, and divide the data point into the cluster represented by the shortest center point; S503: after all data points are divided, update the center points of the p clusters using the mean of the data points in the cluster; S504: repeat S502 and S503 until the center points no longer change, and the training stops; S60: calculating the recommended purchase value of auto parts store goods and the product package attribution according to the integrated prediction model, the safety stock level model and the clustering analysis model; determining the sales prediction value of the store's single product according to the integrated prediction model, outputting the recommended purchase value through the category safety stock level model combined with the current inventory of the store, and outputting the product purchase priority combined with the product package clustering model.

2. The method according to claim 1, wherein, The training data set contains data of multiple commodities, and for any commodity, a data set corresponds ; is a feature data set of the i-th sample, is the next month sales volume of the i-th sample, and the feature data includes past sales volume of a preset number of months, purchase volume, number of sales customers, average purchase price of the current month, number of newly developed customers, single price of repair plant, purchase of active service providers, and purchase of non-active first batch service providers.

Images