A data-driven VMI outbound demand forecasting method in lean manufacturing
Patent Information
- Application Number
- CN202611189730.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-06
- Publication Date
- 2026-10-27
AI Technical Summary
1.解决现有技术忽视上游排产方案对下游物料需求物理驱动效应的问题。本发明引入生产多样性熵作为物理外生约束特征，量化车间精益排产的复杂程度及其对物料波动的边界影响。
Smart Images
Figure CN122988941A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of lean manufacturing technology, and more specifically, to a data-driven method for forecasting VMI outbound demand in lean manufacturing. Background Technology
[0002] Mixed-flow production, characterized by diverse product types and small batches, has become the primary production method for discrete manufacturing enterprises to meet the dynamic and diverse needs of customers. Under this model, high-value manufacturing enterprises face a dual dilemma: the risk of stockouts and inventory backlogs coexist in their Vendor Managed Inventory (VMI) model due to the drastic fluctuations in demand caused by multi-variety, small-batch production. Managers need to predict the material outbound demand generated from production orders in order to allocate resources such as manpower, inventory, and facilities more efficiently and better serve production and delivery.
[0003] However, traditional data-driven forecasting methods often fail to capture the non-stationary demand characteristics in complex mixed-flow scheduling environments, while ignoring the physical constraints of material consumption brought about by production planning. This results in low efficiency of dynamic adjustments in lean supply chains and difficulty in achieving optimal operational levels.
[0004] Existing VMI (Vehicle Management) material requirements forecasting technologies have several technical shortcomings when adapting to lean mixed-flow manufacturing scenarios, failing to simultaneously meet the requirements for forecasting accuracy and inventory risk management. Specific deficiencies are as follows: 1. Poor adaptability to demand patterns and low forecast accuracy. Traditional statistical models such as SARIMA and Crosston assume that time series data is stable and have a weak ability to capture demand peaks and sudden demand, resulting in large forecast errors. They are unable to take into account both continuous and intermittent demand patterns, leading to inventory backlog and stockout risks. They are also unable to adapt to the non-stationary and multi-scale fluctuating demand sequences brought about by multi-model mixed production.
[0005] 2. The loss function is out of touch with industrial realities. Existing models all use symmetric loss functions, which do not reflect the actual cost structure of discrete manufacturing, where "losses from stockouts and shutdowns far exceed the costs of warehousing and inventory backlogs." The penalty weights for overestimating and underestimating demand are the same, which can easily lead to production line shutdowns due to material shortages.
[0006] 3. Lack of production mechanism constraints leads to severe prediction bias. Traditional models rely solely on historical outbound data for modeling, failing to incorporate the internal mechanisms of mixed-flow scheduling in the workshop. They cannot quantify the amplifying and smoothing effects of multi-model mixed scheduling complexity on material demand fluctuations. Without physical mechanism constraints, prediction results are severely biased under extreme scheduling scenarios.
[0007] 4. Significant temporal noise interference, prone to overfitting. The timing of raw material outbound production is superimposed with multiple layers of noise, including weekly production cycles, monthly capacity fluctuations, and urgent orders. Traditional models do not perform multi-scale decoupling processing, and directly fitting the original data easily leads to overfitting and underfitting, resulting in a complete loss of predictive reference value.
[0008] In summary, existing technologies lack an integrated prediction framework that combines production mechanisms, time-frequency domain decoupling, bidirectional time-series feature mining, and asymmetric risk loss. They cannot balance prediction accuracy with inventory risk management and are insufficient to support quantitative decision-making on safety stock under the lean manufacturing VMI model. Summary of the Invention
[0009] The technical problem to be solved by the present invention is to address the above-mentioned shortcomings of the prior art. The purpose of the present invention is to provide a data-driven method for predicting VMI outbound demand in lean manufacturing.
[0010] The technical solution of this invention is: a data-driven method for forecasting VMI outbound demand in lean manufacturing, comprising the following steps S: Step S1. Collect data from the target VMI warehouse within a preset historical time period. Historical outbound data sequence and the corresponding production schedule. ; Step S2. Use variational mode decomposition algorithm to process the outbound data sequence. Decompose into a specified number Eigenmode function components with different center frequencies ; Step S3. According to the production schedule Calculate during the preset historical time period Production diversity entropy ; Step S4. The intrinsic mode function components With production diversity entropy Combine them to form a multidimensional input feature matrix and the multidimensional input feature matrix A multidimensional input tensor is constructed by performing normalization processing, and a sample set is constructed based on the multidimensional input tensor; Step S5. Using the multidimensional input tensor as input and the production scheduling suggestion as output, construct a BiLSTM network, which is composed of a forward LSTM layer and a backward LSTM layer. Step S6. Based on the sample set, and using the Pinball Loss loss function as the loss function of the BiLSTM network, train the model to obtain the outbound demand prediction model; Step S7. Deploy the outbound demand forecasting model to the enterprise VMI management system or MES system, and collect the time period before the current time. The outbound data and production scheduling data are processed according to steps S2 to S4 to construct a multidimensional input tensor, which is then input into the outbound demand prediction model to obtain the time period after the current moment. Recommended production volume within the specified range.
[0011] As a further improvement, the production scheduling plan The data includes the planned production volume of each model per day.
[0012] Furthermore, in step S1, the collected historical outbound data sequence is processed. and production schedule The data is cleaned or missing values are handled.
[0013] Furthermore, in step S2, the variational mode decomposition algorithm is implemented by constructing and solving the following constrained variational problem:
[0014] The constraints are:
[0015] in, The first decomposition obtained Each intrinsic mode function component For the first The center frequency of each component For the Dirac function, This represents the convolution operation. The imaginary unit; Introducing a quadratic penalty term parameter and Lagrange multipliers The constrained variational problem is transformed into an unconstrained generalized Lagrangian function:
[0016] Iterative optimization is performed using the alternating direction method of multipliers to update components in the frequency domain. Center frequency Lagrange multipliers The calculation formulas are as follows:
[0017]
[0018]
[0019] in, The update step size for the multiplier is set, and the update termination condition is set as follows:
[0020] in, To achieve convergence tolerance.
[0021] Furthermore, in step S3, let the engine model be... The total production capacity of the system on that day was:
[0022] Time machine model The production share is:
[0023] Based on Shannon's information entropy principle, define Entropy of production diversity at any moment for:
[0024] in, The boundary constraints must be met.
[0025] The production diversity entropy It has the following boundary characteristics: if only a single model is produced on a given day, then , If all If these models are produced in a proportional mixed-flow manner, then... , Reaching the maximum value .
[0026] Furthermore, in step S4, the normalization formula is: .
[0027] Furthermore, in step S5, LSTM introduces a gating mechanism to control the accumulation and forgetting of information; at a specific time... The input vector is The hidden state at the previous moment was The cell state at the previous moment was The calculation formula for its internal control door is as follows: Forgotten Gate Determines the cell state at the previous moment The proportion of information that needs to be discarded is:
[0028] Input gate With candidate cell state Control the input vector at the current time step The proportion of cell states written is:
[0029]
[0030] Current cell state update for:
[0031] Output gate With the current hidden state for:
[0032]
[0033] in, This is the weight matrix. It is the bias vector; Use the Sigmoid activation function; It is the hyperbolic tangent activation function; This represents the Hadamard product.
[0034] Furthermore, the forward LSTM layer extracts the forward dependency features of the sequence in ascending time order, and the backward LSTM layer extracts the backward dependency features of the sequence in reverse time order. The formula for calculating the hidden state sequence of the feedforward LSTM layer is:
[0035] The formula for calculating the hidden state sequence of the backward LSTM layer is:
[0036] Finally, by concatenating the forward and backward hidden state sequences, a feature vector incorporating bidirectional time-dependent features is output. : .
[0037] Furthermore, the Pinball Loss function is:
[0038]
[0039] in, This is the actual demand value. These are the model's predicted values. For quantile parameters, To predict the unit loss caused by underestimation, To predict the unit loss caused by an overestimation.
[0040] Furthermore, Set it to be no less than 0.9.
[0041] Beneficial effects Compared with the prior art, the present invention has the following advantages: 1. This invention addresses the problem of existing technologies neglecting the physical driving effect of upstream production scheduling on downstream material demand. It introduces production diversity entropy as a physical exogenous constraint feature to quantify the complexity of lean production scheduling in the workshop and its boundary impact on material fluctuations.
[0042] 2. This invention addresses the technical problem that existing prediction models cannot effectively handle the strong non-stationarity and multi-scale noise interference of VMI outbound demand sequences. It employs variational mode decomposition (VMD) to decouple the non-stationary outbound demand sequence in the frequency domain, constructing a combined prediction model based on VMD and a bidirectional long short-term memory network, effectively solving the problems of noise interference and overfitting.
[0043] 3. This invention addresses the problem that existing symmetric loss functions cannot match the cost asymmetry of prediction bias in VMI business scenarios. It introduces an asymmetric pinball loss function for optimization, enabling risk-sensitive dynamic inventory level decisions.
[0044] 4. This invention solves the technical problems of poor fitting of non-stationary mixed demand, homogenization of prediction bias penalties, failure to utilize prior production information, and uncontrollable inventory risk in traditional technologies. At the same time, it takes into account the accuracy of time series fitting and the constraints of stockout risk, and can output quantitative replenishment levels with confidence intervals. This provides a quantitative and scientific basis for decision-making on safety stock configuration, intelligent production scheduling, and supplier collaborative stock preparation under the lean production VMI model. Attached Figure Description
[0045] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0046] The present invention will be further described below with reference to specific embodiments shown in the accompanying drawings.
[0047] See Figure 1 A data-driven method for forecasting VMI outbound demand in lean manufacturing includes the following steps S1 to S7: Step S1. Data Acquisition and Preprocessing Data collection of target VMI warehouses within a preset historical time period Historical outbound data sequence and the corresponding production schedule. .
[0048] Production schedule The data includes the planned production quantity for each model each day, and the collected historical outbound data sequences. and production schedule Data cleaning or missing value handling is performed. Data cleaning includes removing duplicate, erroneous, or outlier values (such as negative numbers), and missing value handling is performed using interpolation or average values.
[0049] Step S2. Frequency domain decoupling is performed on the non-stationary outbound demand sequence using variational mode decomposition (VMD). The variational mode decomposition algorithm is used to analyze the outbound data sequence in step S1. Decompose into a specified number Eigenmode function components with different center frequencies .
[0050] Specifically, the variational mode decomposition algorithm is implemented by constructing and solving the following constrained variational problem:
[0051] The constraints are:
[0052] in, The first decomposition obtained Each intrinsic mode function component For the first The center frequency of each component For the Dirac function, This represents the convolution operation. It is the imaginary unit.
[0053] Introducing a quadratic penalty term parameter and Lagrange multipliers The constrained variational problem is transformed into an unconstrained generalized Lagrangian function:
[0054] Iterative optimization is performed using the Alternating Direction Method of Multipliers (ADMM). Components are updated in the frequency domain. Center frequency Lagrange multipliers The calculation formulas are as follows:
[0055]
[0056]
[0057] in, The update step size for the multiplier is set, and the update termination condition is set as follows:
[0058] in, To achieve convergence tolerance.
[0059] In variational mode decomposition (VMD), the selection of the number of decomposition modes K determines the physical rationality of the decomposition results. If K is too small, it can easily lead to under-decomposition, causing multiple frequency features to overlap in a single component; if K is too large, it can easily produce spurious over-decomposition components. To quantitatively determine the number of decomposition modes K suitable for the dataset used in this invention, it is necessary to perform power spectral density (PSD) analysis on the total daily outbound demand of parts throughout the year, and set the number of decomposition modes K for the VMD algorithm based on the number of dominant frequency components.
[0060] This invention proposes a quantitative determination method based on power spectral density analysis. It calculates the dominant frequency components of the demand sequence using discrete Fourier transform, and uses the number of dominant frequency components plus one as the decomposition mode number. This method ensures that each intrinsic mode function component obtained from the decomposition corresponds to different production management scales, including long-term production baseline trends, work-week production cycles, intra-week production scheduling fine-tuning, short-term Kanban scheduling cycles, and high-frequency random noise. This avoids the subjectivity and blindness of traditional methods that rely on experience or trial-and-error to set the decomposition level.
[0061] Step S3. Calculate the Production Diversity Entropy (PDE) as an exogenous constraint feature. According to the production schedule Calculate within a preset historical time period Production diversity entropy .
[0062] Let the engine model be Seed, within a preset historical time period The production schedule for each model is as follows: The total production capacity of the system on that day is:
[0063] Time machine model The production share is:
[0064] Based on Shannon's information entropy principle, define Entropy of production diversity at any moment for:
[0065] in, The boundary constraints must be met.
[0066] Production diversity entropy It has the following boundary characteristics: if only a single model is produced on a given day, then , If all If these models are produced in a proportional mixed-flow manner, then... , Reaching the maximum value .
[0067] This invention introduces production diversity entropy into the field of supplier managed inventory outbound demand forecasting, using it as a physical characteristic to quantify the complexity of multi-variety, small-batch production scheduling. Through the nonlinear constraint relationship between production diversity entropy and the upper bound of component demand fluctuations, the physical driving mechanism of production scheduling structure on material demand is mathematically proven. When production scheduling tends towards a proportionally mixed flow, production diversity entropy increases, and demand fluctuations are spontaneously smoothed out; when production scheduling is concentrated on a few models, production diversity entropy decreases, and demand fluctuations are amplified.
[0068] Step S4. Construct a multidimensional input tensor and perform normalization. The K eigenmode function components obtained from step S2 are decomposed. With production diversity entropy Combine them to form a multidimensional input feature matrix and the multidimensional input feature matrix Normalization is performed to construct a multidimensional input tensor.
[0069] In this embodiment, the minimum-maximum normalization method is used to eliminate dimensional differences between features. The normalization principle formula is as follows:
[0070] The value is in the range of 0 to 1. express Any one of the features, express The minimum value in, express The maximum value in.
[0071] After normalization, a multidimensional input tensor is constructed as the input to the subsequent BiLSTM network.
[0072] Based on multiple historical time periods The multidimensional input tensor is used to construct the sample set. According to the time order, the sample set is divided into training set, validation set and test set in a ratio of 7:1:2 to ensure the temporal continuity of the data in each set and avoid future information leakage.
[0073] Step S5. Construct and train a Bidirectional Long Short-Term Memory (BiLSTM) network. A BiLSTM network is constructed using a multidimensional input tensor as input and a production scheduling suggestion as output. The BiLSTM network is composed of a forward LSTM layer and a backward LSTM layer.
[0074] LSTM controls the accumulation and forgetting of information by introducing a gating mechanism; at a specific moment... The input vector is The hidden state at the previous moment was The cell state at the previous moment was The calculation formula for its internal control door is as follows: Forget Gate Determines the cell state at the previous moment The proportion of information that needs to be discarded is:
[0075] Input Gate With candidate cell state Control the input vector at the current time step The proportion of cell states written is:
[0076]
[0077] Current cell state update for:
[0078] Output Gate With the current hidden state for:
[0079]
[0080] in, This is the weight matrix. It is the bias vector; Use the Sigmoid activation function; It is the hyperbolic tangent activation function; This represents the Hadamard product.
[0081] The forward LSTM layer extracts the forward dependency features of the sequence in ascending time order, while the backward LSTM layer extracts the backward dependency features of the sequence in reverse time order.
[0082] The formula for calculating the hidden state sequence of the feedforward LSTM layer is:
[0083] The formula for calculating the hidden state sequence of the backward LSTM layer is:
[0084] Finally, by concatenating the forward and backward hidden state sequences, a feature vector incorporating bidirectional time-dependent features is output. :
[0085] In this embodiment, the model parameters are set to a single-layer BiLSTM network with 64 hidden units to balance model complexity and computational efficiency. A Dropout layer is added after the BiLSTM layer with a dropout rate of 0.2 to suppress overfitting. The ReLU activation function is used for the fully connected layers. The training strategy uses the Adam optimization algorithm to update the network weight parameters, with an initial learning rate of 0.001 and a learning rate decay strategy to mitigate gradient fluctuations. The batch size is set to 32, the maximum number of iterations is set to 200, and an early stopping mechanism is introduced: training stops when the validation set loss no longer decreases within 20 consecutive iterations.
[0086] Step S6. Set the Pinball Loss function for asymmetric optimization training. Based on the sample set, and using the Pinball Loss function as the loss function of the BiLSTM network, the model is trained to obtain the outbound demand prediction model.
[0087] Based on the asymmetry between the risk of insufficient production caused by forecasts being lower than actual demand and the risk of excessive production caused by forecasts being higher than actual demand in actual business operations, a quantile parameter τ (0 < τ < 1) is set.
[0088] According to the critical quantile principle of the newsboy model, the quantile parameter τ is equal to the optimal quantile that minimizes the expected total cost of the system, that is, the proportion of the penalty weight for the predicted value being lower than the actual demand to the total penalty weight of the two types of deviations:
[0089] in, To predict the unit loss caused by underestimation, To predict the unit loss caused by an overestimation.
[0090] The Pinball Loss function is:
[0091] in, This is the actual demand value. These are the model's predicted values. This is the quantile parameter.
[0092] In this embodiment, Setting the value to no less than 0.9 ensures that the penalty for underestimating a prediction is nine times that for overestimating it, thus guaranteeing that the model will automatically tend to output higher predictions during training to statistically cover potential demand fluctuations.
[0093] During training, the validation set loss is monitored, and an early stopping mechanism is used to prevent overfitting. The model parameters with the optimal validation set loss are saved.
[0094] This invention replaces the traditional symmetric loss function with Pinball Loss. By leveraging the critical quantile principle of the newsboy model, it directly maps the business cost structure (the ratio of under-predicted loss to over-predicted loss) to the quantile parameter τ, naturally biasing the model output towards covering high-demand scenarios. Specifically, when τ = 0.9, the model output is the 90th quantile of the demand distribution. This value can be directly used as a VMI scheduling recommendation, statistically ensuring that there is only a 10% probability of under-scheduling.
[0095] Step S7. Predict the output production scheduling recommendation quantity Deploy the outbound demand forecasting model to the enterprise's VMI management system or MES system, and collect data from the time period prior to the current moment. The outbound data and production scheduling data are processed according to steps S2 to S4 to construct a multidimensional input tensor, which is then input into the outbound demand prediction model to obtain the time period after the current moment. Recommended production volume within the specified range.
[0096] This invention constructs a prediction architecture that combines variational mode decomposition, a bidirectional long short-term memory network, and a Pinball Loss asymmetric loss function. Variational mode decomposition decouples complex non-stationary signals into several relatively stationary narrowband components to reduce fitting difficulty. The bidirectional long short-term memory network extracts the bidirectional dependency features of each component in its temporal evolution. The Pinball Loss asymmetric loss function enables asymmetric risk preference regulation, thus solving the joint optimization problem of non-stationary signal processing, long-term dependency modeling, and asymmetric risk decision-making.
[0097] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention, and these will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.
Claims
1. A data-driven method for forecasting VMI outbound demand in lean manufacturing, characterized in that, Includes the following steps S: Step S1. Collect data from the target VMI warehouse within a preset historical time period. Historical outbound data sequence and the corresponding production schedule. ; Step S2. Use variational mode decomposition algorithm to process the outbound data sequence. Decompose into a specified number Eigenmode function components with different center frequencies ; Step S3. According to the production schedule Calculate during the preset historical time period Production diversity entropy ; Step S4. The intrinsic mode function components With production diversity entropy Combine them to form a multidimensional input feature matrix and the multidimensional input feature matrix A multidimensional input tensor is constructed by performing normalization processing, and a sample set is constructed based on the multidimensional input tensor; Step S5. Using the multidimensional input tensor as input and the production scheduling suggestion as output, construct a BiLSTM network, which is composed of a forward LSTM layer and a backward LSTM layer. Step S6. Based on the sample set, and using the Pinball Loss loss function as the loss function of the BiLSTM network, train the model to obtain the outbound demand prediction model; Step S7. Deploy the outbound demand forecasting model to the enterprise VMI management system or MES system, and collect the time period before the current time. The outbound data and production scheduling data are processed according to steps S2 to S4 to construct a multidimensional input tensor, which is then input into the outbound demand prediction model to obtain the time period after the current moment. Recommended production volume within the specified range.
2. The data-driven lean manufacturing VMI outbound demand forecasting method according to claim 1, characterized in that, The production schedule The data includes the planned production volume of each model per day.
3. The data-driven lean manufacturing VMI outbound demand forecasting method according to claim 1, characterized in that, In step S1, the collected historical outbound data sequence is processed. and production schedule The data is cleaned or missing values are handled.
4. The data-driven lean manufacturing VMI outbound demand forecasting method according to claim 1, characterized in that, In step S2, the variational mode decomposition algorithm is implemented by constructing and solving the following constrained variational problem: The constraints are: in, The first decomposition obtained Each intrinsic mode function component For the first The center frequency of each component For the Dirac function, This represents the convolution operation. The imaginary unit; Introducing a quadratic penalty term parameter and Lagrange multipliers The constrained variational problem is transformed into an unconstrained generalized Lagrangian function: Iterative optimization is performed using the alternating direction method of multipliers to update components in the frequency domain. Center frequency Lagrange multipliers The calculation formulas are as follows: in, The update step size for the multiplier is set, and the update termination condition is set as follows: in, To achieve convergence tolerance.
5. The data-driven lean manufacturing VMI outbound demand forecasting method according to claim 1, characterized in that, In step S3, let the engine model be... The total production capacity of the system on that day was: Time machine model The production share is: Based on Shannon's information entropy principle, define Entropy of production diversity at any moment for: in, The boundary constraints must be met. The production diversity entropy It has the following boundary characteristics: if only a single model is produced on a given day, then , If all If these models are produced in a mixed-flow manner at a constant ratio, then... , Reaching the maximum value .
6. The data-driven lean manufacturing VMI outbound demand forecasting method according to claim 1, characterized in that, In step S4, the normalization formula is: 。 7. The data-driven lean manufacturing VMI outbound demand forecasting method according to claim 1, characterized in that, In step S5, LSTM introduces a gating mechanism to control the accumulation and forgetting of information; at a specific time... The input vector is The hidden state at the previous moment was The cell state at the previous moment was The calculation formula for its internal control door is as follows: Forgotten Gate Determines the cell state at the previous moment The proportion of information that needs to be discarded is: Input gate With candidate cell state Control the input vector at the current time step The proportion of cell states written is: Current cell state update for: Output gate With the current hidden state for: in, This is the weight matrix. It is the bias vector; Use the Sigmoid activation function; It is the hyperbolic tangent activation function; This represents the Hadamard product.
8. A data-driven method for predicting VMI outbound demand in lean manufacturing according to claim 7, characterized in that, The forward LSTM layer extracts the forward dependency features of the sequence in ascending time order, and the backward LSTM layer extracts the backward dependency features of the sequence in reverse time order. The formula for calculating the hidden state sequence of the feedforward LSTM layer is: The formula for calculating the hidden state sequence of the backward LSTM layer is: Finally, by concatenating the forward and backward hidden state sequences, a feature vector incorporating bidirectional time-dependent features is output. : 。 9. A data-driven method for predicting VMI outbound demand in lean manufacturing according to claim 1, characterized in that, The Pinball Loss function is: in, This is the actual demand value. These are the model's predicted values. For quantile parameters, To predict the unit loss caused by underestimation, To predict the unit loss caused by an overestimation.
10. A data-driven method for predicting VMI outbound demand in lean manufacturing according to claim 9, characterized in that, Set it to be no less than 0.9.