A supply chain demand prediction method based on conditional guided timing diffusion and multi-modal fusion

Abstract

The present application relates to artificial intelligence, in particular to a supply chain demand prediction method based on conditional guided time series diffusion and multi-modal fusion, comprising: obtaining multi-modal business samples containing missing mask historical sales and picture-text description; obtaining cross-modal alignment feature vectors through feature extraction and spatial projection, and constructing static multi-modal condition representation; under the guidance, using a time series diffusion network to perform reverse denoising sampling to obtain reconstructed time series features; generating filling uncertainty measurement signals through multiple sampling variance calculations, inputting dynamic gating network to adaptively allocate fusion weights, and obtaining global cross-modal attribute features; finally, generating demand prediction results, and performing end-to-end training based on the joint loss function of prediction error and diffusion noise prediction loss. The present application effectively overcomes the supply chain cold start and out-of-stock fault problems, and significantly improves the prediction accuracy and model robustness in the extreme sparse scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data processing technology, and in particular to a supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion. Background Technology

[0002] In modern retail and e-commerce operations, supply chain demand forecasting is a crucial issue for enterprises in inventory management and capacity planning. With the accelerating pace of product iteration, accurate and efficient forecasting of future sales is essential for reducing the risk of overstocking and avoiding stockout losses.

[0003] Traditional supply chain demand forecasting methods typically rely on historical sales trends of a single product for sequential modeling. If the model is based solely on data from a single numerical domain, it fails to capture the potential driving force of product visual style and textual attributes on sales. Furthermore, traditional sequential models cannot effectively handle the large number of missing values ​​(masks) caused by stockouts and the cold start problem for new products in real-world scenarios, thus limiting forecast accuracy. Existing deep learning-based multimodal forecasting methods often employ static concatenation during feature fusion, lacking awareness of temporal uncertainty. This indirectly leads to severe feature interference and parameter overfitting issues in extremely sparse or highly volatile data. Summary of the Invention

[0004] To improve the accuracy of forecasting, this invention proposes a supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion, specifically including the following steps:

[0005] S1. Obtain the preprocessed multimodal business sample set of the target product, wherein the multimodal business sample includes: the original historical sales sequence of the target product with the real missing mask, and the corresponding product image data and product text attribute data;

[0006] S2. Perform feature extraction and spatial projection on the multimodal business samples, extract the visual feature vector of the product image data and the text feature vector of the product text attribute data respectively, and project them into the shared semantic space to obtain cross-modal aligned feature vectors;

[0007] S3. Construct a static multimodal conditional representation based on cross-modal aligned feature vectors. Construct a training missing mask for the non-missing time steps with real sales labels in the original historical sales sequence. Input the historical sales sequence with the training missing mask into the temporal diffusion network. Perform inverse denoising sampling under the guidance of cross attention of the static multimodal conditional representation to obtain the reconstructed temporal feature vector.

[0008] S4. Calculate the variance of the multiple sampling distribution of the reconstructed temporal feature vector, extract the variance features of the diffusion filling process, and generate a filling uncertainty measurement signal for the current missing mask.

[0009] S5. Construct a dynamic gating network based on the uncertainty measurement signal, and use the dynamic gating network to assign adaptive fusion weights to the reconstructed temporal feature vector and the static multimodal conditional representation to obtain the global cross-modal attribute features after gating soft fusion.

[0010] S6. Based on the global cross-modal attribute features, the demand forecasting model generates future sales forecast results, and a joint loss function is constructed by combining the prediction error and the diffusion noise prediction loss at the corresponding position of the training missing mask to train the system end-to-end; the trained demand forecasting system is used to predict the demand for the historical sales sequence of the test product with the real missing mask and its multimodal data.

[0011] This invention extracts cross-modal alignment features of images and text through a cross-attention mechanism, comprehensively capturing the potential driving force of product visual style and text attributes on sales, breaking through the representation limitations of traditional single numerical time-series data. Simultaneously, it introduces a time-series diffusion network trained on historical sales training samples to perform inverse denoising and reconstruction of the original sales sequence with missing masks, effectively overcoming the problem of extremely sparse data caused by abnormal stockouts and cold starts of new products in the supply chain, and avoiding statistical biases introduced by traditional interpolation methods. Furthermore, this invention fills in uncertainties through multiple sampling quantization, utilizes a dynamic gating network to adaptively allocate feature fusion weights to isolate high-variance noise interference, and combines an end-to-end joint loss function for global optimization, avoiding parameter overfitting in extreme fluctuation scenarios. Ultimately, it significantly improves prediction accuracy, system robustness, and global decision reliability in complex supply chain environments. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion according to the present invention.

[0013] Figure 2 This is a schematic diagram of the method flow framework of the present invention;

[0014] Figure 3 This is a training diagram of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] This invention proposes a supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion, such as... Figure 1 Specifically, it includes the following steps:

[0017] S1. Obtain the preprocessed multimodal business sample set of the target product, wherein the multimodal business sample includes: the original historical sales sequence of the target product with the real missing mask, and the corresponding product image data and product text attribute data;

[0018] S2. Perform feature extraction and spatial projection on the multimodal business samples, extract the visual feature vector of the product image data and the text feature vector of the product text attribute data respectively, and project them into the shared semantic space to obtain cross-modal aligned feature vectors;

[0019] S3. Construct a static multimodal conditional representation based on cross-modal aligned feature vectors. Construct a training missing mask for the non-missing time steps with real sales labels in the original historical sales sequence. Input the historical sales sequence with the training missing mask into the temporal diffusion network. Perform inverse denoising sampling under the guidance of cross attention of the static multimodal conditional representation to obtain the reconstructed temporal feature vector.

[0020] S4. Calculate the variance of the multiple sampling distribution of the reconstructed temporal feature vector, extract the variance features of the diffusion filling process, and generate a filling uncertainty measurement signal for the current missing mask.

[0021] S5. Construct a dynamic gating network based on the uncertainty measurement signal, and use the dynamic gating network to assign adaptive fusion weights to the reconstructed temporal feature vector and the static multimodal conditional representation to obtain the global cross-modal attribute features after gating soft fusion.

[0022] S6. Based on the global cross-modal attribute features, the demand forecasting model generates future sales forecast results, and a joint loss function is constructed by combining the prediction error and the diffusion noise prediction loss at the corresponding position of the training missing mask to train the system end-to-end; the trained demand forecasting system is used to predict the demand for the historical sales sequence of the test product with the real missing mask and its multimodal data.

[0023] As an optional implementation, the data preprocessing process for the target product multimodal business sample set in this embodiment includes:

[0024] S11: Extract historical transaction logs from the supply chain, aggregate the sales volume and price change characteristics of the target product within a preset time window, and generate an original continuous time series. Preferably, to eliminate high-frequency noise in the underlying transaction logs, this embodiment uses "day" or "week" as the granularity to slide and aggregate the past data of the target product. Based on transaction data at each time step, those skilled in the art can flexibly set time windows according to the life cycle of the target product. Size.

[0025] S12: Based on the system inventory status, identify supply chain out-of-stock intervals, abnormal zero-sales intervals, or new product data-free intervals in the continuous time series, and replace the values ​​at the corresponding time steps with a missing mask identifier (MASK) to obtain the original historical sales sequence with the true missing mask. In this embodiment, the system abandons the traditional mean or zero-filling strategy and forces the true out-of-stock intervals to be set to the MASK state. In particular, if the target product is a new product with zero historical transactions, its sequence is initialized to a length of [length missing]. The full mask sequence preserves the true fault distribution and aims to provide a clear reconstruction target for subsequent temporal diffusion networks.

[0026] S13: Based on the unique identifier (SKU) of the target product, extract the corresponding global visual image from the distributed database, and simultaneously clean and concatenate the discrete text description of the product, ultimately combining them into a structured multimodal business sample. Preferably, the extracted original visual image is cropped and scaled to a uniform resolution matrix, while unstructured character tags in the text description are removed, thereby reducing the computational complexity of feature extraction and ensuring the quality of the input signal for subsequent cross-modal alignment.

[0027] In this embodiment, product image data and product text attribute data are independently feature-encoded. Specifically, the preprocessed product image data is input into a visual feature extraction network, employing a Vision Transformer model based on a self-attention mechanism. Through image patch feature extraction and positional encoding injection, a local visual feature matrix preserving the spatial topology is extracted, denoted as... , This indicates the number of image blocks after image segmentation. The feature dimension is represented by the input of product text attribute data into the BERT text semantic encoding network to extract the attribute intent feature matrix of the text word sequence. , Indicates the number of valid word segments in a text sequence.

[0028] In this embodiment, considering that product images in real e-commerce scenarios may contain redundant information unrelated to the product itself, such as background areas, shooting environment, and display scene, directly using global image features for sales prediction could easily reduce the correlation between visual features and the prediction task. Therefore, as Figure 2 This embodiment introduces a cross-modal attention alignment mechanism, using the text attribute intent feature matrix. As a query condition, the local visual feature matrix Semantic relevance calculation and weighted aggregation are performed to obtain aligned visual features guided by text attributes, specifically including the following steps:

[0029] First, the query matrix is ​​generated through a linear transformation. Key matrix Sum matrix The calculation formula is as follows:

[0030]

[0031]

[0032]

[0033] in, The weight projection matrix is ​​a learnable matrix. The hidden layer dimension after projection;

[0034] Visual feature matrix after cross-modal attention alignment The calculation formula is:

[0035]

[0036] in, A semantic relevance score matrix representing the relationship between text words and various image patches. Scaling factor It is a normalized exponential function; This represents a representation matrix that, after being purified by text attributes, highly focuses on the core visual features of the product.

[0037] To unify the feature dimensions for subsequent use by the temporal diffusion network, the aligned visual feature matrix was... Spatial projection fusion, static multimodal condition representation The calculation formula is:

[0038]

[0039] in, Represents a multilayer perceptron network; This represents a matrix concatenation operation; This indicates that the matrix is ​​subjected to global average pooling along the sequence dimension; Represented by the attribute intent feature matrix The global text feature vector obtained after pooling.

[0040] By performing the aforementioned cross-modal attention alignment calculation on multimodal business samples, this invention enables abstract text attributes to actively find and highlight corresponding local visual regions in the image, automatically filtering complex background noise while spatially projecting, thereby greatly improving the prediction accuracy of downstream demand forecasting systems under complex commodity forms.

[0041] In this embodiment, the original historical sales sequence is input into a temporal diffusion network trained based on historical sales training samples. Inverse denoising sampling is performed under the guidance of cross-attention in static multimodal conditional representation to obtain a reconstructed temporal feature vector, such as... Figure 2 Specifically, it includes:

[0042] The temporal diffusion network trained based on historical sales training samples includes a forward diffusion perturbation stage, and its transition probability distribution is calculated using the following formula:

[0043]

[0044] in, This represents the sequence of latent variables at the t-th diffusion time step, given the sequence of latent variables at the (t-1)-th diffusion time step. The conditional transition probability distribution; Indicates the first The sequence of latent variables after adding noise at each diffusion time step; This represents the sequence of latent variables at the (t-1)th diffusion time step; express Follow the mean The covariance matrix is Gaussian distribution, It is the identity matrix. These are the preset variance scheduling parameters.

[0045] Preferably, inverse denoising sampling is performed under the guidance of cross-attention in static multimodal conditional representation, and the formula for calculating the transition probability distribution of its inverse reconstruction is as follows:

[0046]

[0047] in, Represents a noisy sequence of latent variables at a given diffusion time step t. And in the case of static multimodal conditional representation C, inverse denoising is used to generate the latent variable sequence at the (t-1)th diffusion time step. The conditional transition probability distribution; express Follow the mean The covariance matrix is Gaussian distribution; The learnable parameters of a temporal diffusion network are represented. This indicates that the temporal diffusion network, guided by the static multimodal conditional representation C, according to... The inverse denoising mean obtained by predicting the diffusion time step t, This represents the inverse denoising covariance matrix predicted by the temporal diffusion network under the same conditions.

[0048] In this embodiment, the forward diffusion perturbation process of the temporal diffusion network actually involves gradually injecting Gaussian noise into the complete and unmissing real historical sales sequence of the target product through a Markov chain to simulate the discontinuities and abnormal missing states in the real supply chain scenario. Meanwhile, in the reverse denoising sampling process, the network uses the original historical sales sequence with a real missing mask as a basis, and then applies it to the static multimodal conditional representation vector. Noise is gradually removed under global guidance. By introducing a temporal diffusion network to reconstruct the mask sequence, the problem of extremely sparse data caused by abnormal stockouts and cold starts of new products in the supply chain is effectively overcome. This avoids the statistical bias introduced by traditional interpolation methods and helps to improve the model's generalization ability and reconstruction accuracy in extreme scenarios.

[0049] It should be noted that since the temporal diffusion network has already learned the joint distribution of real data between multimodal image and text attributes and sales sequences during the training phase, there is no need to perform a forward diffusion perturbation stage when forecasting demand for the product under test in practical applications; instead, the missing mask portion in the test sequence is directly initialized as pure Gaussian noise. ,go through Step-by-step iterative noise reduction sampling At any given time, a reconstructed temporal feature vector integrating image and text attribute features is directly output. .

[0050] This embodiment uses a temporal diffusion network to independently perform K inverse denoising sampling operations on the original historical sales sequence with a missing mask to obtain a set of candidate reconstructed sequences. .

[0051] The k-th candidate complete reconstruction sequence is represented as:

[0052]

[0053] in, This represents the Hadamard product operation, where M is the missing mask matrix, and the elements in the matrix are M. ijThis represents the mask value of the i-th target product at the j-th historical time step. A value of 1 indicates that the sales time step has been observed, and a value of 0 indicates that the sales time step is missing. This refers to the observed portion of the original historical sales sequence; This represents the reconstruction result of the missing position obtained from the k-th inverse denoising sampling.

[0054] Calculate the distribution variance at each time step to obtain the uncertainty metric matrix U, which is expressed as:

[0055]

[0056] in, This represents the Hadamard product operation. The mean of the complete reconstructed sequence obtained from K samplings represents the overall historical sales estimate of the time-series diffusion network after filling in the missing sales figures. In this embodiment, the filling uncertainty metric matrix U is used to quantify the filling uncertainty of the model at the missing time steps.

[0057] In this embodiment, the reconstructed sequence of the target product is sampled multiple times and the distribution variance is calculated. This allows for precise quantification of the diffusion model's "lack of confidence" in filling stockout gaps. Smaller variance indicates greater confidence in the model's ability to fill sales gaps during those periods; larger variance indicates significant uncertainty regarding the outcome. For example, for a popular product in its normal sales lifecycle that only experiences occasional stockouts of one or two days, the diffusion network, based on its rich temporal dependencies, demonstrates highly consistent reconstruction results with each sampling, resulting in an uncertainty metric matrix. The value is extremely small; however, for a newly launched product, due to the lack of historical references, the results of each sampling by the diffusion network can vary drastically, leading to a calculated uncertainty metric matrix. The value is extremely large.

[0058] It should be noted that traditional deep learning prediction models typically only output a deterministic point estimate and cannot assess the confidence level of their own prediction results; however, this invention, by introducing multiple sampling variance, endows the model with the ability to self-evaluate, providing a crucial quantitative indicator for subsequent adaptive noise isolation features.

[0059] This embodiment will fill in the uncertainty metric matrix. The input is fed into a gated perception layer consisting of a multilayer perceptron and activation functions to generate an adaptive cross-modal fusion weight vector. , is represented as:

[0060]

[0061] in, and These are the trainable weight matrix and bias vector of the gated perception layer, respectively; To map the output to The normalized activation function for the interval.

[0062] Furthermore, the mean sequence Mapped to deep dynamic temporal features via one-dimensional convolutional layers ; Utilizing the fusion weight vector ,right With static multimodal conditional representation vector By performing feature interactions based on Hadamard product, global cross-modal attribute features are obtained, namely:

[0063]

[0064] In this embodiment, when the filled variance of a certain dimension When the value is too large, after passing through the Sigmoid activation function with a negative weight parameter matrix, the corresponding weights... It automatically decays and approaches 0. Through this soft fusion mechanism, the system adaptively cuts off unreliable temporal reconstruction feature paths and instead relies on static multimodal conditional representation vectors. This method enables prediction, effectively isolating noise interference caused by high reconstruction variance and preventing the model from overfitting in extreme fluctuation scenarios.

[0065] For example, the variance of the best-selling product in step S4. The weights output by the gating network are extremely small at this point. Approaching 1, the system mainly relies on deep dynamic temporal characteristics. To make accurate predictions; and for the first launch of a new product, its variance great, As the value decays to near zero, the system dynamically transfers the prediction weights to... On one hand, it relies entirely on the static images and textual attributes of the initial product launch. To infer initial sales volume.

[0066] It is worth noting that in practical applications, the fusion ratio of traditional feature concatenation methods is fixed after model training; however, due to the presence of a dynamic gating network, the system can adapt to changes in real time for each product under test during the testing phase. Even for the same product, the extracted fusion weights remain consistent across different supply and demand phases. It is also completely different, greatly improving the predictive robustness in the broader retail scenario.

[0067] During end-to-end training, this invention incorporates global cross-modal attribute features. Input a fully connected prediction layer, output a future preset time window Forecast sales Calculate and predict sales With real labels Demand forecasting loss between , is represented as:

[0068]

[0069] Where H is the prediction time step; To represent the future number Step-by-step sales forecast This represents the corresponding actual sales volume.

[0070] Simultaneously, during the forward and backward processes of the temporal diffusion network, the diffusion noise prediction loss between the network's predicted noise and the actual injected Gaussian noise is calculated. When the original historical sales sequence If the data at a certain time step is complete, the value of the diffusion noise prediction loss is 0; otherwise, the diffusion noise prediction loss at that time step is zero. Represented as:

[0071]

[0072] in, This represents the Gaussian noise injected into the original historical sales sequence x0. And the expected value of the diffusion time step t; This indicates the calculation of the L2 norm; Gaussian noise injected into reality, Representation of temporal diffusion networks under static multimodal conditions Noise is used for conditionally guided prediction.

[0073] The demand forecasting loss and the diffusion noise forecasting loss are linearly combined to construct an end-to-end joint loss function. , is represented as:

[0074]

[0075] in, , To balance the weight hyperparameters of the prediction and reconstruction tasks.

[0076] In this embodiment, traditional multimodal prediction frameworks often treat "data imputation" and "downstream prediction" as two separate stages for training, resulting in the imputed data not necessarily being the most suitable for the prediction task. For example... Figure 3 This invention constructs a joint loss function that includes prediction error and spread noise prediction loss. The backpropagation algorithm is used to perform end-to-end joint parameter optimization on the entire network. This mechanism can learn the intrinsic relationship between temporal features and image attributes with extremely stable performance, so that the reconstruction process of the upstream diffusion network is directly controlled by the downstream prediction accuracy target, thereby achieving the synergistic effect of "intelligently filling data and accurately predicting demand" and further significantly improving the accuracy of demand forecasting in complex supply chain scenarios.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion, characterized in that, Specifically, the following steps are included: S1. Obtain the preprocessed multimodal business sample set of the target product, wherein the multimodal business sample includes: the original historical sales sequence of the target product with the real missing mask, and the corresponding product image data and product text attribute data; S2. Perform feature extraction and spatial projection on the multimodal business samples, extract the visual feature vector of the product image data and the text feature vector of the product text attribute data respectively, and project them into the shared semantic space to obtain cross-modal aligned feature vectors; S3. Construct a static multimodal conditional representation based on cross-modal aligned feature vectors. Construct a training missing mask for the non-missing time steps with real sales labels in the original historical sales sequence. Input the historical sales sequence with the training missing mask into the temporal diffusion network. Perform inverse denoising sampling under the guidance of cross attention of the static multimodal conditional representation to obtain the reconstructed temporal feature vector. S4. Use a temporal diffusion network to independently perform K inverse denoising samplings. The reconstructed temporal feature vectors from the K inverse denoising samplings constitute a candidate reconstruction sequence set. Calculate the variance of the candidate reconstruction sequence set at the corresponding missing time step to obtain the filling uncertainty measurement signal U for the current missing mask. S5. Construct a dynamic gating network based on the uncertainty-filling metric signal. Utilize this dynamic gating network to assign adaptive fusion weights to the reconstructed temporal feature vector and the static multimodal conditional representation, obtaining the global cross-modal attribute features after gated soft fusion, including: in, Global cross-modal attribute features; For adaptive fusion weights; This represents the Hadamard product operation; For deep dynamic temporal features, the mean of the sequence is obtained by sampling the reconstructed temporal feature vector multiple times and then mapping it through a one-dimensional convolutional layer; This is a static multimodal condition representation; Represents the Sigmoid function; Let be the trainable weight matrix in the gated sensing layer after non-negativity constraints. This is the bias vector of the gated perception layer; S6. Based on the global cross-modal attribute features, the demand forecasting model generates future sales forecast results, and a joint loss function is constructed by combining the prediction error and the diffusion noise prediction loss at the corresponding position of the training missing mask to train the system end-to-end; the trained demand forecasting system is used to predict the demand for the historical sales sequence of the test product with the real missing mask and its multimodal data.

2. The supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion according to claim 1, characterized in that, The data preprocessing process for the target product multimodal business sample set includes: Extract historical transaction logs from the supply chain, and statistically aggregate the sales data of the target product within each time window according to the preset time window to generate the original sales time series arranged in chronological order. By combining the inventory status, shelf time, sales status and order fulfillment records of the target product, identify the supply chain out-of-stock intervals, abnormal zero sales intervals or new product no historical data intervals in the original continuous sales time series, and set missing mask identifiers for the corresponding time steps to obtain the original historical sales series with the true missing mask. Based on the unique identifier of the target product, the corresponding product image data is obtained, and the text attribute data of the target product is cleaned and spliced. Finally, the original historical sales sequence with real missing mask, product image data and product text attribute data are combined into a structured multimodal business sample.

3. The supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion according to claim 1, characterized in that, The process involves extracting visual feature vectors from product image data and text feature vectors from product text attribute data, then projecting these two vectors into a shared semantic space to obtain cross-modal aligned feature vectors. The specific steps include: The product image data is input into a visual feature extraction network to obtain a local visual feature matrix in the spatial dimension, denoted as F. image After the product text attribute data is processed by word segmentation, vectorization, and positional encoding, it is input into a text semantic coding network to obtain the attribute intent sequence matrix, denoted as F. text ; Based on F text Mapping to obtain the query vector, based on F image The mapping yields key and value vectors. Cross-modal cross-attention computation is then performed based on the query vector, key vector, and value vector to obtain the aligned visual feature matrix. Global pooling is performed on the aligned visual feature matrix along the spatial dimension to obtain a global visual feature vector; pooling is performed on the attribute intent sequence matrix along the text sequence dimension to obtain a global text feature vector; the global visual feature vector and the global text feature vector are mapped to preset feature dimensions respectively and then concatenated, and mapped to a unified implicit semantic space through a multilayer perceptron to obtain a cross-modal aligned feature vector.

4. The supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion according to claim 1, characterized in that, Performing inverse denoising sampling guided by cross-attention in static multimodal conditional representation includes the following steps: During the training phase, training samples are constructed based on the historical sales sequence of the target product, and a training missing mask is constructed based on the non-missing time steps with real sales labels. Gaussian noise is gradually injected into the time steps corresponding to the training missing mask through a forward diffusion process to obtain noisy sales sequences under different diffusion time steps. The cross-modal aligned feature vector is mapped to a static multimodal conditional representation vector through a multilayer perceptron network. At each inverse denoising time step, the current temporal hidden state of the temporal diffusion network is mapped to a query matrix, and the static multimodal conditional representation vector is mapped to a key matrix and a value matrix. Cross-attention calculation is performed based on the query matrix, key matrix, and value matrix to obtain conditional guided features. After fusing the conditional guided features into the current temporal hidden state, the noise corresponding to the current inverse denoising time step is predicted, and the noisy sales sequence is updated accordingly. During the reconstruction phase, the missing time steps in the original historical sales sequence with the true missing mask are initialized as Gaussian noise, and stepwise inverse denoising sampling is performed under the guidance of conditional guided features. In each inverse denoising time step, only the missing time steps are denoised and updated, while the sales information of the observed time steps remains unchanged. Finally, the reconstructed temporal feature vector that integrates product image data and product text attribute data is output.

5. The supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion according to claim 1, characterized in that, The uncertainty measure signal U for filling the current missing mask is expressed as: Where M is the missing mask matrix, and the elements in the matrix are M ij This represents the mask value of the i-th target product at the j-th historical time step. A value of 1 indicates that the sales time step has been observed, and a value of 0 indicates that the sales time step is missing. This represents the Hadamard product operation. This represents the mean of the complete reconstructed sequence obtained from K samplings. Characterize the time-series diffusion network for estimating overall historical sales after missing sales figures are filled in.

6. The supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion according to claim 1, characterized in that, The loss function for end-to-end training is expressed as: in, The loss function for end-to-end training; To predict losses; Predicting loss for diffused noise; , To balance the weight hyperparameters of the prediction and reconstruction tasks.

7. The supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion according to claim 6, characterized in that, Predicting losses Represented as: in, To predict the time step; Indicates the future number Forecasted sales for each step; For the future The actual sales figures for the product.

8. The supply chain demand forecasting method based on condition-guided temporal diffusion and multimodal fusion according to claim 6, characterized in that, When the original historical sales sequence If the data at a certain time step is complete, the value of the diffusion noise prediction loss is 0; otherwise, the diffusion noise prediction loss at that time step is zero. Represented as: in, This indicates the original historical sales sequence. Real-injected Gaussian noise And the expected value of the diffusion time step t; This indicates the calculation of the L2 norm; Gaussian noise injected into reality, Representation of temporal diffusion networks under static multimodal conditions As noise in conditionally guided prediction; Indicates the original historical sales sequence The noisy sales sequence is obtained by adding noise at the t-th diffusion time step.

Images