CN110837959A - Method for balancing and fixing welding dispatching working hours based on welding quantity and operation mode - Google Patents

Abstract

The invention provides a method for balancing and determining welding dispatching man-hour based on the quantity of welding materials and an operation mode, which comprises the following steps: step S1) data acquisition; step S2) data integration; step S3), data screening, and selecting the most relevant important dimension of the target task; step S4), data preprocessing including data specification and data transformation; step S5), data mining is carried out by adopting a multilayer feedforward neural network, and a welding man-hour prediction model is established; step S6) inputting the quantity and the operation mode into the prediction model to obtain the working hours of the quantity to be processed; the invention has the advantages that: the accurate working hours of welding each work order are accurately predicted to be an entry point, the working hours of each welding work order are accurately predicted, and the fine parts are accurately quantized as much as possible, so that the overall required working hours of macroscopic accurate prediction are achieved, and the aims of accurately predicting the large construction period and the delivery period are finally achieved.

Description

Method for balancing and fixing welding dispatching working hours based on welding quantity and operation mode

Technical Field

The invention relates to the field of ship manufacturing, in particular to a method for balancing and fixing welding dispatching work hours based on the quantity of welding materials and an operation mode.

Background

The labor management of ship enterprises belongs to production management, the current on-site ship building mode of China is generally an outsourced mode, and a generation management department of the ship manufacturing enterprises issues work instructions to each team and then carries out specific dispatching by team leaders. However, in some domestic ship building enterprises, due to unreasonable production arrangement or quality problems of labor teams, the phenomena of waiting for work and idling often occur, the occurrence of the phenomena greatly affects the accuracy of a production period, the production efficiency and the enthusiasm of staff, and the research on how to avoid the problems has good research significance.

The field of production dispatching in the ship manufacturing production process still belongs to the traditional dispatching mode. The traditional welding man-hour prediction often has great deviation, can not consider comprehensively and meticulously, divides the task to the operation area one-level, carries out the work arrangement by the team leader, and the subjective consciousness and the randomness, the tendency of team leader can play more main effect, and this can lead to unfair phenomenon and the emergence of time limit delay, advance scheduling problem to a certain extent.

Disclosure of Invention

The invention aims to provide a method for balancing and fixing the work hours for dispatching welding based on the quantity of welding materials and the operation mode.

In order to achieve the purpose, the technical scheme of the invention is as follows:

a method for balancing and determining welding dispatching man-hour based on the quantity of welding materials and an operation mode is characterized by comprising the following steps:

step S1), collecting operation data of the same work type in the intelligent ship subsection workshop system, and recording subsection welding processing production data;

step S2), integrating the heterogeneous and time-series production data together, and integrating the data in a plurality of dispersed data sources into a data set according to the time series or the classification of samples and the logic or the physics of different data;

step S3), data screening is carried out on the welding data, and the most relevant important dimension for the target task is selected;

step S4), preprocessing the screened data, wherein the preprocessing comprises data specification and data transformation;

step S5), data mining is carried out by adopting a multilayer feedforward neural network, and a welding man-hour prediction model is established;

step S6) inputs the amount and the operation mode into the prediction model, and obtains the man-hour for which the amount needs to be processed.

Further, in step S3, the most relevant important dimensions are the factors that most affect the man-hours, including the welding method, the welding form, the type of the part, whether the sheet is a straight line sheet, the plate thickness, the length of the weld, whether the sheet is a straight line, the welding position, the working stage, and the assembling flow direction.

Further, in step S5, the data mining of the multi-layer feedforward neural network includes the following steps:

step S501) initializing a network;

step S502) taking a sample, and inputting the sample into the model in a forward direction;

step S503) calculating neuron input and output of each layer;

step S504) calculating the error between the output value and the actual value;

step S505) the error is reversely propagated, the partial derivative is calculated through the weight of the transmission path between different neurons, and the weight and the threshold are adjusted;

step S506) judging whether the learning samples are used up, if so, skipping to the step S507, otherwise, returning to the step S502;

step S507) calculating the average error E of the network;

step S508) judging whether the precision of the average error E meets the requirement, if so, ending the learning process, otherwise, skipping to the step S509;

step S509), determining whether the iteration number reaches an upper limit, if so, ending the learning process, otherwise, returning to step S506.

Further, the establishing of the welding man-hour prediction model includes: the historical data is learned through the neural network, potential corresponding relations are found out, the corresponding relations between input dimensions and output man-hours are mined, and a multilayer feedforward neural network model with fixed weight is formed.

Further, the forming of the multilayer feedforward type neural network model with fixed weight value comprises the following steps:

step S511) data preprocessing;

step S512), establishing a neural network structure;

step S513) judging whether the training reaches the set times, if so, skipping to step S516, and if not, skipping to step S514;

step S514) selecting data of a batch from the training data set;

step S515) training neural network parameters, and returning to step S513;

step S516), calculating the error of the test data set;

step S517) judging whether the training frequency reaches an upper limit, if so, modifying the neural network structure, returning to the step S513, if not, continuing to judge whether the error meets the requirement, if so, jumping to the step S518, and if not, returning to the step S513;

step S518) predicts unknown data.

The invention has the advantages that: the method includes the steps that specific welding working hour values are predicted by a plurality of welding influence factors, the corresponding relation between important dimensions and welding working hour values is found through a machine learning method, after rules between the important dimensions are found, test data are predicted, the accuracy of the test data is compared, the accurate working hour of each work order is accurately predicted to be an entry point, the working hour of each welding work order is accurately predicted, the fine parts are quantized accurately as much as possible, the overall required working hour of macroscopic accurate prediction is achieved, and the purposes of accurately predicting a large construction period and a delivery period are achieved finally.

Drawings

FIG. 1 is a flow chart of data mining in accordance with the present invention;

FIG. 2 is a schematic diagram of the learning process of the multi-layer feedforward neural network of the present invention;

FIG. 3 is a network structure diagram of a simple multi-layer feedforward neural network used in the present invention;

FIG. 4 is a flow chart of data mining with mass balance timing according to the present invention;

FIG. 5 is a diagram illustrating the effect of the model of the present invention on predicting unknown data after learning;

FIG. 6 is a graph showing the effect of learning rate degradation after the exponential decay learning rate method is adopted in the present invention.

Detailed Description

The technical solutions of the present invention will be described clearly and completely with reference to the accompanying drawings, and it should be understood that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The embodiment discloses a method for balancing and determining welding dispatching man-hour based on the amount of welding materials and an operation mode, as shown in fig. 1, comprising the following steps:

step S1), collecting operation data of the same work type in the intelligent ship segmented workshop system, and recording segmented welding processing production data, preferably collecting and recording the production data from the welding dimensions as much as possible based on a system with higher informatization degree such as MES and the like;

step S2), integrating the heterogeneous and time-series production data together, and integrating the data in a plurality of dispersed data sources into a data set according to the time series or the classification of samples and the logic or the physics of different data;

step S3), data screening is carried out on the welding data, and the most relevant important dimension for the target task is selected;

step S4), preprocessing the screened data, wherein the preprocessing comprises data specification and data transformation, and a data set which is approximately equivalent to or even better than the original data set but has less data volume is obtained through the data specification; changing data from one expression form to another expression form through data transformation, wherein the common data transformation modes are data standardization, discretization and semantic conversion;

step S5), data mining is carried out by adopting a multilayer feedforward neural network, and a welding man-hour prediction model is established;

step S6) inputs the amount and the operation mode into the prediction model, and obtains the man-hour for which the amount needs to be processed.

In step S3, the most relevant important dimensions are the factors that most affect the man-hours, including the welding method, the welding form, the type of the part, whether the sheet is a straight line sheet, the sheet thickness, the length of the weld, whether the sheet is a straight line, whether the sheet is a machined panel, the welding position, the working stage, and the assembly flow direction.

In step S5, a multi-layer feedforward neural network is used for data mining, and the neural network has the following advantages: 1) any function can be approximated with any precision; 2) the neural network method belongs to a nonlinear model and can adapt to various complex data relationships; 3) the neural network has strong learning ability, so that the neural network can better adapt to the change of a data space than a plurality of classification algorithms; 4) the neural network can simulate certain functions of the human brain by using the physical structure and mechanism of the human brain for reference, and has the characteristic of intelligence.

The learning process of the multilayer feedforward neural network consists of two processes of forward propagation of signals and backward propagation of errors, the signals calculate weighted sums of all layers from an input layer during forward propagation, the weighted sums are finally transmitted to an output layer through all hidden layers to obtain output results, and the output results are compared with expected results to obtain output errors; the error back propagation is to carry out back propagation layer by layer from the hidden layer to the input layer according to a gradient descent algorithm, distribute the error to all units of each layer, thereby obtaining an error signal of each unit, and modify the weight of each unit according to the error signal; the two signal processes are continuously circulated to update the weight value, and whether the circulation is finished or not is judged according to the judgment condition.

The multilayer feedforward type neural network preferably adopts a single-layer hidden layer network, wherein the single-layer hidden layer network comprises an input layer, a hidden layer and an output layer, the activation function of the multilayer feedforward type neural network is required to be differentiable, a sigmoid function can be generally adopted as the activation function, and the sigmoid function can be adopted by both the hidden layer and the output layer.

The output at the hidden layer is represented as:

the output at the output layer is represented as:

describing the network error by using a cost function E, adjusting the parameters in the negative gradient direction of the cost function by using a random gradient descent algorithm, and updating the weight value only aiming at one training sample each time, wherein the algorithm is called as an error back propagation algorithm.

And (3) reversely adjusting the weight value by gradient descent, wherein the error descends fastest along the gradient direction according to a gradient descent strategy, so that the adjustment amount of the weight value is in direct proportion to the gradient descent of the error, namely:

for the output layer:

for hidden layers:

in the formula, η ∈ (0,1) indicates a learning rate, and is used to limit the training speed.

The output error for the network is defined as:

for the pth sample:

for all samples:

in specific implementation, as shown in fig. 2, in step S5, the data mining of the multi-layer feedforward neural network includes the following steps:

step S501) initializing a network;

step S502) taking a sample, and inputting the sample into the model in a forward direction;

step S503) calculating neuron input and output of each layer;

step S504) calculating the error between the output value and the actual value;

step S505) the error is reversely propagated, the partial derivative is calculated through the weight of the transmission path between different neurons, and the weight and the threshold are adjusted;

step S506) judging whether the learning samples are used up, if so, skipping to the step S507, otherwise, returning to the step S502;

step S507) calculating the average error E of the network;

step S508) judging whether the precision of the average error E meets the requirement, if so, ending the learning process, otherwise, skipping to the step S509;

step S509), determining whether the iteration number reaches an upper limit, if so, ending the learning process to obtain an optimal solution, otherwise, returning to step S506.

As shown in fig. 3, it is a network structure diagram of a multilayer feedforward type neural network; learning historical data through a neural network, finding out a potential corresponding relation, mining the corresponding relation between input dimensionality and output man-hour, forming a multilayer feedforward neural network model with fixed weight as shown in FIG. 4, wherein the forming of the multilayer feedforward neural network model with fixed weight specifically comprises the following steps:

step S511) data preprocessing;

step S512), establishing a neural network structure;

step S513) judging whether the training reaches the set times, if so, skipping to step S516, and if not, skipping to step S514;

step S514) selecting data of a batch from the training data set;

step S515) training neural network parameters, and returning to step S513;

step S516), calculating the error of the test data set;

step S517) judging whether the training frequency reaches an upper limit, if so, modifying the neural network structure, returning to the step S513, if not, continuing to judge whether the error meets the requirement, if so, jumping to the step S518, and if not, returning to the step S513;

step S518) predicts unknown data.

In the field of welding dispatching in the ship manufacturing production process, a large number of welding dispatching list records are studied, and the actual relation between the important welding dimension and the welding working hour is excavated, wherein the relation is not only established on the basis of the quantity of materials, but also implicitly covers the selection of the efficiency of personnel, the quality of equipment and materials.

This test was performed by sampling 800 pieces of history records, in which 90% of the data was learned and the other 10% of the data was predicted from the learned model, and the prediction result was shown in fig. 5, and in 10% of the data that was not seen in the machine, the RMSE by hand:0.0626936109362 (root mean square error: 0.06269) between the prediction result and the actual welding man-hour was obtained.

On the basis, the working hours can be predicted more quickly by adopting an exponential decay learning rate method; as shown in fig. 6, after a certain number of iterations, the learning rate (i.e., learning step size) has decayed to 20% of the original, and fine tuning is performed instead of coarse-type tuning parameters.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims (5)

1. A method for balancing and determining welding dispatching man-hour based on the quantity of welding materials and an operation mode is characterized by comprising the following steps:

step S1), collecting operation data of the same work type in the intelligent ship subsection workshop system, and recording subsection welding processing production data;

step S2), integrating the heterogeneous and time-series production data together, and integrating the data in a plurality of dispersed data sources into a data set according to the time series or the classification of samples and the logic or the physics of different data;

step S3), data screening is carried out on the welding data, and the most relevant important dimension for the target task is selected;

step S4), preprocessing the screened data, wherein the preprocessing comprises data specification and data transformation;

step S5), data mining is carried out by adopting a multilayer feedforward neural network, and a welding man-hour prediction model is established;

step S6) inputs the amount and the operation mode into the prediction model, and obtains the man-hour for which the amount needs to be processed.

2. The method for determining the labor hour for welding based on the amount of the welding object and the operation manner as claimed in claim 1, wherein in step S3, the most relevant important dimension is the most influencing factor of the labor hour, and the most influencing factor of the labor hour comprises the welding method, the welding form, the part type, whether the assembly line sheet is straight, the plate thickness, the length of the welding seam, whether the panel is linear, whether the panel is processed, the welding position, the operation stage and the assembly flow direction.

3. The method for balancing the labor hour for dispatching welding based on the quantity of welding objects and the operation mode as claimed in claim 1, wherein in step S5, the data mining of the multi-layer feedforward type neural network comprises the following steps:

step S501) initializing a network;

step S502) taking a sample, and inputting the sample into the model in a forward direction;

step S503) calculating neuron input and output of each layer;

step S504) calculating the error between the output value and the actual value;

step S505) the error is reversely propagated, the partial derivative is calculated through the weight of the transmission path between different neurons, and the weight and the threshold are adjusted;

step S506) judging whether the learning samples are used up, if so, skipping to the step S507, otherwise, returning to the step S502;

step S507) calculating the average error E of the network;

step S508) judging whether the precision of the average error E meets the requirement, if so, ending the learning process, otherwise, skipping to the step S509;

step S509), determining whether the iteration number reaches an upper limit, if so, ending the learning process, otherwise, returning to step S506.

4. The method for determining the labor hour for dispatching welding based on the quantity of welding objects and the operation mode according to claim 1, wherein the establishing the welding labor hour prediction model comprises: the historical data is learned through the neural network, potential corresponding relations are found out, the corresponding relations between input dimensions and output man-hours are mined, and a multilayer feedforward neural network model with fixed weight is formed.

5. The method for determining the dispatching time of the welding based on the quantity of the welding objects and the operation mode according to claim 4, wherein the step of forming a multilayer feedforward type neural network model with fixed weight value comprises the following steps:

step S511) data preprocessing;

step S512), establishing a neural network structure;

step S513) judging whether the training reaches the set times, if so, skipping to step S516, and if not, skipping to step S514;

step S514) selecting data of a batch from the training data set;

step S515) training neural network parameters, and returning to step S513;

step S516), calculating the error of the test data set;

step S517) judging whether the training frequency reaches an upper limit, if so, modifying the neural network structure, returning to the step S513, if not, continuing to judge whether the error meets the requirement, if so, jumping to the step S518, and if not, returning to the step S513;

step S518) predicts unknown data.

Images