Manufacturing logistics intelligent matching recommendation method and system based on multi-modal data

Abstract

The invention provides a manufacturing logistics intelligent matching recommendation method and system based on multi-modal data, and relates to the technical field of intelligent logistics, and the method comprises the steps: obtaining the logistics demand data of a manufacturing enterprise and the service capability data of a logistics supplier; constructing a multi-level feature attention network to perform feature extraction on the logistics demand data and the service capability data, and performing screening to obtain multi-modal feature representation; constructing a positive and negative sample pair based on the multi-modal feature representation, obtaining an inter-modal semantic mapping matrix through a hierarchical contrast learning mode, and projecting the multi-modal feature representation to a unified semantic space to obtain a fusion feature representation; the fusion feature representation and the historical cooperation record are constructed into a state vector, an instant reward function and a long-term reward function are constructed, and a matching degree score between the manufacturing enterprise and the logistics supplier is obtained through optimization of a deep reinforcement learning algorithm; and recommending a logistics supplier to the manufacturing enterprise based on the matching degree score.

Description

TECHNICAL FIELD

[0001] The present application relates to intelligent logistics technology, and in particular to a manufacturing logistics intelligent matching recommendation method and system based on multi-modal data. BACKGROUND

[0002] Manufacturing enterprises have increasingly diverse and refined demands for logistics services. Traditional logistics matching methods mainly rely on manual experience or simple rules, which are difficult to adapt to complex and changing market environments. Existing technologies usually use single modal data for matching, which cannot fully capture the multi-dimensional feature information of manufacturing enterprises and logistics suppliers, resulting in low matching accuracy and low efficiency. In particular, when considering factors such as enterprise size, industry attributes, and business characteristics, traditional methods cannot effectively integrate and weigh the influence of these factors, resulting in inaccurate matching, resource waste, and other problems.

[0003] Existing intelligent matching technologies mostly use shallow feature extraction and simple similarity calculation, ignoring the time series, multi-modal, and correlation characteristics of logistics field data. These technologies often focus on short-term matching results and lack consideration of long-term cooperation stability, making them unable to meet the needs of long-term collaborative development between manufacturing enterprises and logistics suppliers. At the same time, existing technologies have semantic gap problems when dealing with heterogeneous data, making it difficult to effectively integrate and map different modal data, limiting the accuracy and personalization of matching recommendations, and making it difficult to meet the requirements of efficient and accurate logistics resource allocation in modern intelligent manufacturing environments. SUMMARY

[0004] To address the deficiencies of existing technologies, the present application provides a manufacturing logistics intelligent matching recommendation method and system based on multi-modal data, which can solve the problems in existing technologies.

[0005] In a first aspect of an embodiment of the present application, a manufacturing logistics intelligent matching recommendation method based on multi-modal data is provided, comprising: obtaining logistics demand data of a manufacturing enterprise and service capability data of a logistics supplier; constructing a multi-level feature attention network to extract features from the logistics demand data and the service capability data, the multi-level feature attention network including an enterprise size layer, an industry attribute layer, and a business feature layer; training the multi-level feature attention network based on historical matching data to obtain feature importance scores, filtering features according to the feature importance scores, and obtaining multi-modal feature representations; constructing a positive and negative sample pair based on the multi-modal feature representations, obtaining an inter-modal semantic mapping matrix through a hierarchical contrast learning method, and projecting the multi-modal feature representations to a unified semantic space using the semantic mapping matrix to obtain a fused feature representation; The fused feature representation and historical cooperation records are constructed into a state vector; an immediate reward function and a long-term reward function are constructed based on the state vector, and a matching score between the manufacturing enterprise and the logistics supplier is obtained through optimization using a deep reinforcement learning algorithm; The logistics orders are divided into time windows and optimally allocated based on the matching scores to obtain an optimized allocation result; and logistics suppliers are recommended to the manufacturing enterprise based on the optimized allocation result.

[0006] Optional, Constructing a multi-level feature attention network to extract features from the logistics demand data and the service capability data, wherein the multi-level feature attention network includes an enterprise scale layer, an industry attribute layer, and a business feature layer; training the multi-level feature attention network based on historical matching data to obtain feature importance scores, screening features according to the feature importance scores, and obtaining a multimodal feature representation includes the following steps: The logistics demand data and the service capability data are constructed into enterprise scale layer features, industry attribute layer features, and business feature layer features according to time series characteristics; in the multi-level feature attention network, a scale layer attention calculation unit is constructed based on the enterprise scale layer features, and a feature stability matrix is ​​introduced to calculate the enterprise scale layer feature weights; an industry layer attention calculation unit is constructed based on the industry attribute layer features, and an industry correlation strength matrix is ​​introduced to calculate the industry attribute layer feature weights; a business layer attention calculation unit is constructed based on the business feature layer features, and a time decay factor is introduced to calculate the business feature layer feature weights; Calculating inter-layer downward enhancement features and inter-layer upward feedback features based on the enterprise scale layer feature weight, the industry attribute layer feature weight, and the business feature layer feature weight; performing weighted fusion on the inter-layer downward enhancement features and the inter-layer upward feedback features to obtain inter-layer interaction features; A multi-objective loss function is constructed based on historical matching data to train the multi-level feature attention network, wherein the multi-objective loss function includes a matching accuracy loss term, a feature stability loss term, and a temporal consistency loss term; the feature independent importance score and the feature interaction importance score are calculated, and the feature importance score is obtained based on the feature independent importance score and the feature interaction importance score; the inter-layer interaction features are screened according to the feature importance score to obtain a multimodal feature representation.

[0007] Optional, The steps of calculating inter-layer downward enhancement features and inter-layer upward feedback features based on the enterprise scale layer feature weight, the industry attribute layer feature weight, and the business feature layer feature weight; and weightedly fusing the inter-layer downward enhancement features and the inter-layer upward feedback features to obtain inter-layer interaction features include: Multiplying the matrices of the enterprise scale layer feature weights, the industry attribute layer feature weights, and the business feature layer feature weights to obtain an inter-layer interaction matrix; calculating the complementarity scores between the feature representations corresponding to the enterprise scale layer feature weights, the industry attribute layer feature weights, and the business feature layer feature weights, and calculating the inter-layer interaction strength based on the complementarity scores and the inter-layer interaction matrix; Constructing an information flow gating unit, which calculates the information flow coefficient based on the characteristic hidden state of adjacent layers to control the transmission strength of information between layers; The inter-layer interaction intensity, the information flow coefficient, the enterprise scale layer feature weight, the industry attribute layer feature weight, and the business feature layer feature weight are combined to calculate the inter-layer downward enhancement feature and the inter-layer upward feedback feature, wherein the inter-layer downward enhancement feature is obtained by the enhancement effect of the high-level feature on the low-level feature, and the inter-layer upward feedback feature is obtained by the feedback effect of the low-level feature on the high-level feature; A feature fusion function is constructed, a dynamic fusion weight is calculated by the feature fusion function, and inter-layer downward enhancement features and inter-layer upward feedback features are weightedly fused according to the dynamic fusion weight to obtain inter-layer interaction features.

[0008] Optional, The steps of constructing positive and negative sample pairs based on the multimodal feature representation, obtaining an inter-modal semantic mapping matrix through a hierarchical contrastive learning method, and projecting the multimodal feature representation into a unified semantic space using the semantic mapping matrix to obtain a fused feature representation include: Constructing a positive sample pair set and a negative sample pair set based on the multimodal feature representation, wherein the positive sample pair set includes matching demand-side feature and supply-side feature pairs, and the negative sample pair set includes mismatched demand-side feature and supply-side feature pairs; Constructing a feature relevance evaluation function that simultaneously considers feature space similarity and industry collaboration frequency; and calculating a dynamic difficulty weight for a sample pair based on the feature relevance evaluation function. Constructing a hierarchical contrastive learning loss function, wherein the hierarchical contrastive learning loss function includes multiple feature levels, each feature level is assigned a corresponding similarity metric function and a temperature parameter; weighting the positive and negative sample pairs of each feature level according to the dynamic difficulty weight to obtain a weighted hierarchical contrastive learning loss; and dynamically adjusting the weight coefficients of different feature levels based on gradient updates; The semantic mapping matrix is ​​optimized based on the hierarchical contrastive learning loss function, and the multimodal feature representation is projected into a unified semantic space through the semantic mapping matrix to obtain a fused feature representation.

[0009] Optional, The steps of constructing the fused feature representation and historical cooperation records into a state vector; constructing an immediate reward function and a long-term reward function based on the state vector, and optimizing the deep reinforcement learning algorithm to obtain a matching score between the manufacturing enterprise and the logistics supplier include: The historical collaboration records are obtained through a temporal attention mechanism, and the temporal attention mechanism calculates a temporal attention weight based on the correlation between the historical collaboration records and temporal context features; Construct a multidimensional reward function, which includes a service matching reward, a cost-effectiveness reward, and a stability reward; the service matching reward is calculated based on the fusion feature similarity and historical service quality scores, the cost-effectiveness reward includes the deviation between the predicted cost and the expected cost, and the stability reward is calculated based on the cooperation volatility and the historical cooperation duration; calculate the immediate reward value based on the multidimensional reward function, and calculate the long-term reward value through temporal discount accumulation; Constructing a dual value network architecture of a state value network and an advantage function network, and calculating the Q value based on the state value output by the state value network and the action advantage output by the advantage function network; Constructing a multi-objective optimization function, the multi-objective optimization function including value loss, policy loss, and entropy loss, wherein the value loss is calculated based on the deviation between the long-term reward value and the state value, the policy loss is calculated based on the action probability distribution and the advantage function, and the entropy loss is used to adjust policy exploration; optimizing the network parameters of the dual-value network architecture based on the multi-objective optimization function; The matching score between the enterprise and the logistics supplier is calculated based on the state value and Q value output by the optimized state value network.

[0010] Optional, The steps to construct a multi-dimensional reward function include: The basic service matching degree is obtained based on the fusion feature similarity, and the service matching degree reward is obtained by combining the historical service quality impact factor based on the time-series decay weight and the adaptive scale factor based on the current business scale; The cost deviation assessment value is calculated based on the deviation between the predicted cost and the expected cost, and combined with the comprehensive benefit index based on the historical logistics efficiency improvement rate and cost saving rate and the market environment factor based on the regional logistics price index to obtain the cost-effectiveness reward; The volatility of cooperation is calculated based on historical cooperation indicators, and the impact factor based on the duration of cooperation and the seasonal adjustment factor based on cyclical fluctuations are combined to obtain stability rewards; The weight coefficients of rewards in each dimension are dynamically adjusted through gradient updates. The rewards in the three dimensions are weighted and combined based on the dynamic weights to obtain the immediate reward value. A time domain adjustment coefficient is introduced to obtain the long-term reward value through temporal discount accumulation.

[0011] Optional, The steps of dividing the logistics orders into time windows and optimizing the allocation based on the matching scores to obtain the optimized allocation results include: Obtain a set of orders to be assigned and calculate the order density within each time period. The order density includes the number of orders per unit time, the uniformity of order distribution, and the order resource occupancy rate. A piecewise function is used to set the time window benchmark size based on the order density. Calculating the urgency of each order based on the order delivery deadline, order priority, and associated order constraints; adaptively adjusting the time window benchmark size based on the urgency, where a high urgency corresponds to a time window contraction factor and a low urgency corresponds to a time window expansion factor; Within the adjusted time window, order allocation constraints are set based on the matching score, including time window constraints, capacity constraints, and order association constraints; the integer programming method is used to optimize the allocation of logistics orders to obtain the optimized allocation results.

[0012] Secondly, we provide a manufacturing logistics intelligent matching recommendation system based on multimodal data, including: The first unit is used to obtain logistics demand data of manufacturing enterprises and service capacity data of logistics suppliers; The second unit is configured to construct a multi-level feature attention network to extract features from the logistics demand data and the service capability data, wherein the multi-level feature attention network includes an enterprise scale layer, an industry attribute layer, and a business feature layer; the multi-level feature attention network is trained based on historical matching data to obtain feature importance scores, and features are screened according to the feature importance scores to obtain a multimodal feature representation; The third unit is used to construct positive and negative sample pairs based on the multimodal feature representation, obtain an inter-modal semantic mapping matrix through a hierarchical contrastive learning method, and use the semantic mapping matrix to project the multimodal feature representation into a unified semantic space to obtain a fused feature representation; A fourth unit is configured to construct a state vector from the fused feature representation and historical cooperation records; construct an immediate reward function and a long-term reward function based on the state vector, and obtain a matching score between the manufacturing enterprise and the logistics supplier through deep reinforcement learning algorithm optimization; The fifth unit is used to divide the logistics orders into time windows and optimize the allocation based on the matching scores to obtain an optimized allocation result; and recommend logistics suppliers to the manufacturing enterprise based on the optimized allocation result.

[0013] According to a third aspect, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0014] By constructing a multi-layered feature attention network, this method effectively captures key information across enterprise scale, industry attributes, and business characteristics, enabling deep feature extraction and optimized screening of logistics demand data and service capability data. This method transcends the limitations of traditional single-modality matching, comprehensively perceiving the multi-dimensional characteristics of manufacturing companies and logistics suppliers, significantly improving matching accuracy. This method is particularly effective in processing large-scale, diverse logistics orders, effectively identifying core influencing factors and filtering out redundant information, significantly improving matching efficiency.

[0015] The present invention solves the semantic gap problem in multimodal data fusion through an innovative architecture that combines hierarchical contrastive learning and deep reinforcement learning, and realizes effective mapping and fusion between different modal data. This method not only focuses on the immediate matching effect, but also considers the stability of cooperation by constructing a long-term reward function, balancing short-term benefits and long-term collaborative development needs. The optimization allocation mechanism based on the time window further improves the utilization efficiency of logistics resources and reduces distribution costs and time consumption. While ensuring the matching accuracy, this method effectively improves the personalization and adaptability of recommendations, and builds an efficient, stable and mutually beneficial collaborative relationship for manufacturing companies and logistics suppliers. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Schematic diagram of the process of an intelligent matching recommendation method for manufacturing logistics based on multimodal data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present invention will be described below in conjunction with the drawings in the embodiments of the present invention. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0018] Figure 1 Schematic diagram of the process of the manufacturing logistics intelligent matching recommendation method based on multimodal data of the present invention, such as Figure 1 As shown, the method includes: Obtain logistics demand data of manufacturing enterprises and service capability data of logistics suppliers; Constructing a multi-level feature attention network to extract features from the logistics demand data and the service capability data, wherein the multi-level feature attention network includes an enterprise scale layer, an industry attribute layer, and a business feature layer; training the multi-level feature attention network based on historical matching data to obtain feature importance scores, and screening features based on the feature importance scores to obtain a multimodal feature representation; constructing positive and negative sample pairs based on the multi-modal feature representation, obtaining an inter-modal semantic mapping matrix through a hierarchical contrastive learning manner; projecting the multi-modal feature representation to a unified semantic space to obtain a fused feature representation by using the semantic mapping matrix; constructing a state vector based on the fused feature representation and the historical cooperation record; constructing an immediate reward function and a long-term reward function based on the state vector, and optimizing to obtain a matching degree score between the manufacturing enterprise and the logistics supplier through a deep reinforcement learning algorithm; dividing and optimizing the allocation of the logistics order based on the matching degree score to obtain an optimized allocation result; and recommending a logistics supplier to the manufacturing enterprise according to the optimized allocation result.

[0019] Optionally, constructing a multi-level feature attention network to extract features from the logistics demand data and the service capability data, the multi-level feature attention network including an enterprise size layer, an industry attribute layer, and a business feature layer; training the multi-level feature attention network based on historical matching data to obtain a feature importance score, and screening features according to the feature importance score to obtain a multi-modal feature representation, the steps comprising: constructing enterprise size layer features, industry attribute layer features, and business feature layer features from the logistics demand data and the service capability data according to time sequence characteristics; in the multi-level feature attention network, constructing a size layer attention calculation unit based on the enterprise size layer features, introducing a feature stability matrix to calculate enterprise size layer feature weights; constructing an industry layer attention calculation unit based on the industry attribute layer features, introducing an industry correlation strength matrix to calculate industry attribute layer feature weights; constructing a business layer attention calculation unit based on the business feature layer features, and introducing a time decay factor to calculate business feature layer feature weights; calculating inter-layer downward enhancement features and inter-layer upward feedback features based on the enterprise size layer feature weights, the industry attribute layer feature weights, and the business feature layer feature weights; and weighting and fusing the inter-layer downward enhancement features and the inter-layer upward feedback features to obtain inter-layer interaction features; training the multi-level feature attention network based on a multi-objective loss function constructed from historical matching data, the multi-objective loss function including a matching accuracy loss term, a feature stability loss term, and a time sequence consistency loss term; calculating a feature independent importance score and a feature interaction importance score, and obtaining a feature importance score based on the feature independent importance score and the feature interaction importance score; screening the inter-layer interaction features according to the feature importance score to obtain a multi-modal feature representation.

[0020] For example, logistics demand data includes information such as the manufacturer's transportation time requirements, cargo type, transportation distance, cargo volume, and cargo weight. Service capability data includes information such as the logistics provider's transportation capacity, vehicle type, service scope, historical service rating, and professional qualifications. For example, the logistics demand data of a manufacturer includes: transportation time requirement of within 3 days, cargo type of electronic products, transportation distance of 500 kilometers, cargo volume of 10 cubic meters, and cargo weight of 2 tons. The service capability data of a logistics provider includes: daily transportation capacity of 50 tons, vehicle types including cold chain vehicles and ordinary trucks, service scope of East China, historical service rating of 4.8 points (out of 5 points), and qualification for dangerous goods transportation.

[0021] Multi-level features are constructed based on time series characteristics. Enterprise scale features include indicators such as annual turnover, number of employees, and asset size. For example, a manufacturing enterprise's turnover over the past three months was RMB 5 million, RMB 5.5 million, and RMB 5.8 million, respectively. It has 200 employees and an asset size of RMB 30 million. Industry attribute features include industry category, upstream and downstream industry correlation, and industry prosperity index. For example, a manufacturing enterprise in the electronic equipment manufacturing industry has an upstream correlation of 0.8 with the semiconductor industry, a downstream correlation of 0.6 with the retail industry, and an industry prosperity index of 125. Business feature features include logistics order frequency, average order value, and timeliness requirements. For example, a manufacturing enterprise has an average monthly logistics order volume of 60 orders, an average order value of RMB 5,000, and an average delivery time requirement of 48 hours.

[0022] When constructing a multi-level feature attention network, we first design three levels of attention calculation units. For the enterprise scale layer, a feature stability matrix is ​​introduced to calculate feature weights. The feature stability matrix is ​​constructed by calculating the rate of change of features within a continuous time window. Features with a rate of change below 10% are assigned a stability value of 0.9, those with a rate of change between 10% and 30% are assigned a stability value of 0.7, and those with a rate of change above 30% are assigned a stability value of 0.5. For a manufacturing enterprise, for example, its turnover changed by 16% over three months, with a stability value of 0.7; its employee number changed by 5%, with a stability value of 0.9; and its asset size changed by 2%, with a stability value of 0.9. By weighting the feature values ​​with the stability values, the feature weights for the enterprise scale layer are: 0.65 for turnover, 0.85 for employee number, and 0.88 for asset size.

[0023] For the industry attribute layer, the industry correlation strength matrix is introduced to calculate the feature weight. The industry correlation strength matrix is constructed based on the national economic industry classification standard and the industry chain correlation degree. The correlation degree is divided into three levels: high, medium and low, corresponding to correlation strength values of 0.9, 0.6 and 0.3, respectively. For example, the correlation strength between the electronic equipment manufacturing industry and the semiconductor industry is 0.9, the correlation strength with the steel industry is 0.5, and the correlation strength with the pharmaceutical industry is 0.2. Through the weighted calculation of the correlation strength and the industry characteristics, the industry attribute layer feature weights are obtained as follows: industry category 0.82, upstream and downstream correlation degree 0.78, and industry sentiment index 0.65.

[0024] For the business feature layer, a time decay factor is introduced to calculate the feature weight. The time decay factor is based on the timeliness of the feature. The decay factor for data within the last 1 month is 0.95, the decay factor for data 1-3 months ago is 0.8, the decay factor for data 3-6 months ago is 0.6, and the decay factor for data more than 6 months ago is 0.4. For example, the weighted value of the order frequency of a manufacturing enterprise in the last 1 month is 57x0.95=54.15, and the weighted value of the order frequency 1-3 months ago is 52x0.8=41.6. Through time decay weighted calculation, the business feature layer feature weights are obtained as follows: order frequency 0.76, single average amount 0.82, and timeliness requirement 0.88.

[0025] Based on the feature weights of the three levels, the inter-layer interaction features are calculated. The inter-layer downward enhancement feature is realized by weighting the lower layer feature with the upper layer feature weight. For example, the feature weight of the turnover of the enterprise size layer 0.65 has a 0.65x0.3=0.195 enhancement effect on the industry category feature of the industry attribute layer. The inter-layer upward feedback feature is realized by feedback adjustment of the upper layer feature with the lower layer feature weight. For example, the feature weight of the order frequency of the business feature layer 0.76 has a 0.76x0.25=0.19 feedback adjustment effect on the industry sentiment index feature of the industry attribute layer. The inter-layer downward enhancement feature and the upward feedback feature are weighted and fused to obtain the inter-layer interaction feature.

[0026] A multi-objective loss function was constructed based on historical matching data to train the network. Mini-batch stochastic gradient descent was used, with an initial learning rate of 0.001. The Adam optimizer was used for parameter updates, with a batch size of 64 and 100 training epochs. To prevent overfitting, L2 regularization was introduced with a coefficient of 0.0001. An early stopping strategy was employed, stopping training when the validation set loss showed no significant decrease for five consecutive epochs. An exponentially decaying learning rate was used during training, with a decay rate of 0.95 every 10 epochs. Historical matching data included previously successfully matched manufacturing company and logistics supplier pairs. For example, an electronic equipment manufacturer and a logistics supplier had successfully collaborated 38 times in the previous 12 months, with an average satisfaction rating of 4.6. The multi-objective loss function consists of three parts: the matching accuracy loss term is calculated by calculating the squared difference between the predicted matching score and the actual matching score. For example, if the predicted matching score is 0.85 and the actual matching score is 0.92, the loss value is 0.0049. The feature stability loss term is calculated by the standard deviation of the feature weight over three consecutive time windows. The smaller the standard deviation, the lower the loss. For example, if the feature weight changes from [0.75, 0.73, 0.77] and the standard deviation is 0.02, the corresponding loss value is 0.04. The temporal consistency loss term is calculated by the deviation between the rate of change of the feature at adjacent time points and the historical average rate of change. For example, if the current rate of change of a feature is 8% and the historical average rate of change is 5%, the deviation is 3%, and the corresponding loss value is 0.09. The final multi-objective loss function value is obtained by setting the matching accuracy weight to 0.5, the feature stability weight to 0.3, and the temporal consistency weight to 0.2, and performing a weighted sum of these three loss terms.

[0027] After training is complete, feature importance scores are calculated. Feature independent importance scores are calculated using the feature permutation method, where a feature is randomly permuted to observe how model performance changes. For example, if model performance decreases by 15% after randomly permuting the turnover feature, its independent importance score is 0.15. Feature interaction importance scores are calculated using the feature combination permutation method, where a feature combination is randomly permuted to observe how model performance changes. For example, if model performance decreases by 25% after randomly permuting the turnover and employee number features, its interaction importance score is 0.25. Based on the independent and interaction importance scores, the feature synergistic importance scores are calculated. For example, the synergistic importance score for the turnover feature is 0.15 + (0.25 - 0.15 - 0.08) × 0.5 = 0.19, where 0.08 is the independent importance score for the employee number feature. Finally, inter-layer interaction features are filtered based on the feature importance scores. A feature filtering threshold is set to 0.1, retaining features with importance scores greater than the threshold and removing features with importance scores less than the threshold. Through screening, we finally get a multimodal feature representation, including turnover and asset size characteristics at the enterprise scale layer, industry category and upstream and downstream correlation characteristics at the industry attribute layer, and order frequency and timeliness requirement characteristics at the business feature layer.

[0028] This paper uses a multi-level feature attention network to extract and optimize features from logistics demand data and service capability data, addressing the feature redundancy and information loss issues faced by traditional methods when processing multimodal data. By introducing a feature stability matrix, an industry correlation strength matrix, and a time decay factor, it achieves differentiated processing of features at different levels. The design of a multi-objective loss function and a feature importance assessment mechanism further optimize the feature screening process and improve the accuracy and efficiency of subsequent matching recommendations.

[0029] Optional, The steps of calculating inter-layer downward enhancement features and inter-layer upward feedback features based on the enterprise scale layer feature weight, the industry attribute layer feature weight, and the business feature layer feature weight; and weightedly fusing the inter-layer downward enhancement features and the inter-layer upward feedback features to obtain inter-layer interaction features include: Multiplying the matrices of the enterprise scale layer feature weights, the industry attribute layer feature weights, and the business feature layer feature weights to obtain an inter-layer interaction matrix; calculating the complementarity scores between the feature representations corresponding to the enterprise scale layer feature weights, the industry attribute layer feature weights, and the business feature layer feature weights, and calculating the inter-layer interaction strength based on the complementarity scores and the inter-layer interaction matrix; Constructing an information flow gating unit, which calculates the information flow coefficient based on the characteristic hidden state of adjacent layers to control the transmission strength of information between layers; The inter-layer interaction intensity, the information flow coefficient, the enterprise scale layer feature weight, the industry attribute layer feature weight, and the business feature layer feature weight are combined to calculate the inter-layer downward enhancement feature and the inter-layer upward feedback feature, wherein the inter-layer downward enhancement feature is obtained by the enhancement effect of the high-level feature on the low-level feature, and the inter-layer upward feedback feature is obtained by the feedback effect of the low-level feature on the high-level feature; A feature fusion function is constructed, a dynamic fusion weight is calculated by the feature fusion function, and inter-layer downward enhancement features and inter-layer upward feedback features are weightedly fused according to the dynamic fusion weight to obtain inter-layer interaction features.

[0030] For example, matrix multiplication is performed between the feature weights of the enterprise scale layer, the industry attribute layer, and the business feature layer to obtain the inter-layer interaction matrix. For a manufacturing enterprise, its enterprise scale layer includes three features: annual turnover, number of employees, and asset size, with corresponding feature weights of 0.65, 0.85, and 0.88, respectively. The industry attribute layer includes three features: industry category, upstream and downstream correlation, and industry prosperity index, with corresponding feature weights of 0.82, 0.78, and 0.65, respectively. The business feature layer includes three features: order frequency, average order value, and timeliness requirements, with corresponding feature weights of 0.76, 0.82, and 0.88, respectively. Multiply the enterprise scale layer feature weight matrix with the industry attribute layer feature weight matrix to obtain the scale-industry interaction matrix. For example, the interaction value between turnover and industry category in the matrix is ​​0.65×0.82=0.533. Similarly, calculate the interaction matrix between the industry attribute layer and the business feature layer to obtain the industry-business interaction matrix. For example, the interaction value between industry category and order frequency is 0.82×0.76=0.623.

[0031] Compare the complementarity scores between feature representations at each level. The complementarity score is measured by the information gain between feature representations; higher information gain indicates stronger complementarity between the features. For each feature representation, use it individually and in combination for prediction, and compare the improvement in predictive performance. For example, if the prediction accuracy of using the enterprise scale layer feature alone is 75%, the prediction accuracy of using the industry attribute layer feature alone is 70%, and the prediction accuracy of using both layers together is 85%, then the complementarity score between the two layers is 85% - 75% - 70% + 60% = 0%, where 60% is the baseline accuracy. In practice, the complementarity score between the scale and industry layer features of a manufacturing enterprise is 0.12, the complementarity score between the industry and business layer features is 0.15, and the complementarity score between the scale and business layer features is 0.08.

[0032] Inter-layer interaction strength is calculated by taking the weighted sum of the elements in the interaction matrix and the corresponding complementarity scores. For example, the interaction strength between the scale and industry layers is calculated by multiplying each element in the scale-industry interaction matrix by the complementarity score of 0.12 and then summing the results to obtain a scale-industry interaction strength of 1.82. Similarly, the industry-business interaction strength is calculated to be 2.01.

[0033] An information flow gating unit is constructed to control the strength of information transmission between layers. The information flow gating unit calculates the information flow coefficient based on the hidden states of features at adjacent layers. For each feature, its importance at the current layer and its correlation with adjacent layers are calculated to generate a gating signal. Taking the annual turnover feature at the scale layer as an example, its importance at the scale layer is 0.65, and its correlation with the industry category feature at the industry layer is 0.533. The calculated gating signal is 0.65 × 0.533 = 0.346. In the gating unit, a sigmoid function is used to map the gating signal to the range of 0-1 to obtain the information flow coefficient. For example, a gating signal of 0.346 for the annual turnover feature, after sigmoid mapping, results in an information flow coefficient of 0.586, indicating that 58.6% of the information in the annual turnover feature can be transferred to the industry layer. For the three characteristics of the enterprise scale layer, the information flow coefficients are calculated to be 0.586, 0.625, and 0.648 respectively; for the three characteristics of the industry attribute layer, the information flow coefficients are calculated to be 0.612, 0.598, and 0.563 respectively.

[0034] By combining the inter-layer interaction strength, information flow coefficient, and feature weights at each layer, we calculate inter-layer downward reinforcement and inter-layer upward feedback features. Inter-layer downward reinforcement is achieved by the reinforcement of lower-layer features by higher-level features. For example, the reinforcement of the scale layer on the industry layer by the annual turnover feature is calculated as follows: the annual turnover feature weight (0.65) multiplied by the information flow coefficient (0.586) multiplied by the scale-industry interaction strength (1.82), resulting in an enhancement value of 0.697. Similarly, the reinforcement of other scale layer features on the industry layer features is calculated, and all reinforcement values ​​for the same target feature are summed to obtain the final downward reinforcement feature value. For example, the total reinforcement value of the scale layer on the industry category feature is 2.15.

[0035] Inter-layer upward feedback features are generated through the feedback from lower-level features to higher-level features. For example, the feedback from the business layer to the industry layer is calculated as follows: the order frequency feature weight (0.76) multiplied by the information flow coefficient (0.612) multiplied by the business-industry interaction intensity (2.01), resulting in a feedback value of 0.935. Similarly, the feedback values ​​of other business layer features on the industry layer are calculated, and all feedback values ​​for the same target feature are summed to obtain the final upward feedback feature value. For example, the total feedback value of the business layer on the industry prosperity index feature is 2.63.

[0036] The feature fusion function uses a weighted average to fuse inter-layer features. Original importance scores are calculated based on the historical matching accuracy of the features. The Softmax mechanism is used to convert these importance scores into initial weights, ensuring that the weights sum to 1 and that the values ​​are reasonably distributed. An exponential smoothing mechanism is then introduced to dynamically fuse the currently calculated initial weights with the historical weights, resulting in dynamic weights with time-series smoothing properties. The smoothing coefficient is used to balance the importance of current and historical information. Finally, the resulting dynamic weights are used to weightedly sum the downward enhancement features and upward feedback features to obtain the final inter-layer interaction features. For example, the downward enhancement feature value is 2.15, and the upward feedback feature value is 1.87. Based on historical matching accuracy performance, the above fusion function calculates a dynamic weight of 0.58 for the downward enhancement feature and 0.42 for the upward feedback feature. The final weighted fusion yields an inter-layer interaction feature value of 2.03 for the industry category features.

[0037] The above method is used to calculate inter-layer interaction characteristics for features at all levels. For a complete feature set from a manufacturing enterprise, for example, the inter-layer interaction calculations for scale-level features yielded interaction characteristics of: annual turnover 1.95, number of employees 2.12, and asset size 2.26; for industry-level features, the inter-layer interaction calculations yielded interaction characteristics of: industry category 2.03, upstream-downstream correlation 1.89, and industry prosperity index 1.75; and for business-level features, the inter-layer interaction calculations yielded interaction characteristics of: order frequency 1.84, average order value 1.97, and timeliness requirement 2.15. These inter-layer interaction characteristics will be used in subsequent feature importance assessment and screening.

[0038] This paper not only captures the interactions between features at different levels but also controls the precision of information flow through a gating mechanism, preventing interference from irrelevant information. The introduction of dynamic fusion weights enables adaptive adjustment of the feature fusion process based on specific scenarios, improving the expressive power of feature representation. The calculation of inter-layer interactive features enhances the semantic richness and discriminative power of features, effectively improving the accuracy and personalization of matching recommendations.

[0039] Optional, The steps of constructing positive and negative sample pairs based on the multimodal feature representation, obtaining an inter-modal semantic mapping matrix through a hierarchical contrastive learning method, and projecting the multimodal feature representation into a unified semantic space using the semantic mapping matrix to obtain a fused feature representation include: Constructing a positive sample pair set and a negative sample pair set based on the multimodal feature representation, wherein the positive sample pair set includes matching demand-side feature and supply-side feature pairs, and the negative sample pair set includes mismatched demand-side feature and supply-side feature pairs; Constructing a feature relevance evaluation function that simultaneously considers feature space similarity and industry collaboration frequency; and calculating a dynamic difficulty weight for a sample pair based on the feature relevance evaluation function. Constructing a hierarchical contrastive learning loss function, wherein the hierarchical contrastive learning loss function includes multiple feature levels, each feature level is assigned a corresponding similarity metric function and a temperature parameter; weighting the positive and negative sample pairs of each feature level according to the dynamic difficulty weight to obtain a weighted hierarchical contrastive learning loss; and dynamically adjusting the weight coefficients of different feature levels based on gradient updates; The semantic mapping matrix is ​​optimized based on the hierarchical contrastive learning loss function, and the multimodal feature representation is projected into a unified semantic space through the semantic mapping matrix to obtain a fused feature representation.

[0040] For example, the set of positive sample pairs includes the feature representations of manufacturers and logistics providers that have been successfully matched historically. For example, if electronic equipment manufacturer A and logistics provider B have successfully collaborated 25 times in the past year, with an average satisfaction rating of 4.7 out of 5, then the multimodal feature representation of manufacturer A and the multimodal feature representation of logistics provider B constitute a positive sample pair. Specifically, the multimodal feature representation of manufacturing enterprise A includes: enterprise scale layer characteristics [annual turnover 1.95, number of employees 2.12, asset scale 2.26], industry attribute layer characteristics [industry category 2.03, upstream and downstream correlation 1.89, industry prosperity index 1.75], and business feature layer characteristics [order frequency 1.84, average order amount 1.97, timeliness requirement 2.15]; the multimodal feature representation of logistics supplier B includes: enterprise scale layer characteristics [transportation capacity 1.82, number of vehicles 2.05, storage area 1.68], industry attribute layer characteristics [service industry coverage 2.21, regional coverage 1.92, degree of specialization 1.85], and business feature layer characteristics [service timeliness 2.13, price level 1.76, service quality 2.32].

[0041] The negative sample pair set includes the feature representations of manufacturing companies and logistics providers that have historically failed or poorly matched. For example, if manufacturing company A and logistics provider C have collaborated three times, but the satisfaction rating is only 2.1, then the feature representation of manufacturing company A and the feature representation of logistics provider C constitute a negative sample pair. The multimodal feature representation of logistics provider C includes: company scale features [capacity 1.45, number of vehicles 1.62, storage area 1.33], industry attribute features [service industry coverage 1.54, regional coverage 1.87, degree of specialization 1.21], and business feature features [service timeliness 1.65, price level 2.11, service quality 1.43].

[0042] The feature correlation evaluation function considers both feature space similarity and industry collaboration frequency. Feature space similarity is calculated using cosine similarity. For example, the feature space similarity between manufacturing company A and logistics supplier B is 0.83 at the enterprise scale level, 0.78 at the industry attribute level, and 0.91 at the business feature level. Industry collaboration frequency is obtained through historical transaction data. For example, the historical collaboration frequency between the electronic equipment manufacturing industry and the logistics supplier is 0.75 (representing 75% of all collaborations). Based on feature space similarity and industry collaboration frequency, a comprehensive correlation score is calculated. For example, the comprehensive correlation score between company A and supplier B is 0.83 × 0.4 + 0.78 × 0.3 + 0.91 × 0.3 + 0.75 × 0.2 = 0.724.

[0043] The dynamic difficulty weight of a sample pair is calculated based on the feature correlation evaluation function. For a positive sample pair, if the comprehensive correlation score is low, it means that the sample pair is difficult to distinguish and should be given a higher weight; for a negative sample pair, if the comprehensive correlation score is high, it means that the sample pair is difficult to distinguish and should also be given a higher weight. Specifically, the dynamic difficulty weight of a positive sample pair is calculated as 1 minus the comprehensive correlation score multiplied by the adjustment factor. For example, the dynamic difficulty weight of Enterprise A and Supplier B is (1-0.724)×1.2=0.331; the dynamic difficulty weight of a negative sample pair is calculated as the comprehensive correlation score multiplied by the adjustment factor. For example, the dynamic difficulty weight of Enterprise A and Supplier C is 0.513×1.2=0.616.

[0044] Construct a hierarchical contrastive learning loss function, which contains multiple feature levels, and each feature level sets a corresponding similarity measurement function and temperature parameter. In this embodiment, three feature levels are set: enterprise scale layer, industry attribute layer, and business feature layer. The similarity measurement function of the enterprise scale layer uses Euclidean distance, and the temperature parameter is set to 0.1; the similarity measurement function of the industry attribute layer uses cosine similarity, and the temperature parameter is set to 0.07; the similarity measurement function of the business feature layer uses dot product, and the temperature parameter is set to 0.05. The temperature parameter controls the smoothness of the similarity distribution. A smaller temperature parameter makes the similarity distribution steeper, which is conducive to distinguishing similar samples from dissimilar samples.

[0045] Positive and negative sample pairs at each feature level are weighted according to the dynamic difficulty weight, resulting in a weighted hierarchical contrastive learning loss. Taking the enterprise scale level as an example, the weighted loss for positive samples between enterprise A and supplier B is the original loss multiplied by the dynamic difficulty weight of 0.331; the weighted loss for negative samples between enterprise A and supplier C is the original loss multiplied by the dynamic difficulty weight of 0.616. The weighted loss values ​​for the three feature levels are: 0.247 for the enterprise scale level, 0.352 for the industry attribute level, and 0.413 for the business feature level. The weight coefficients of different feature levels are dynamically adjusted based on gradient updates, with the initial weight coefficients set to an equal value of 1 / 3. After 10 rounds of training, the weight coefficients of the three levels are adjusted to: 0.28 for the enterprise scale level, 0.32 for the industry attribute level, and 0.40 for the business feature level.

[0046] The semantic mapping matrix is ​​optimized based on a hierarchical contrastive learning loss function. The semantic mapping matrix is ​​a multidimensional matrix used to map feature representations from different modalities into a unified semantic space. Taking the mapping matrix from manufacturing enterprise features to the unified semantic space as an example, a 9×6 matrix is ​​initially randomly generated (9 represents the original feature dimension, and 6 represents the unified semantic space dimension). Similarly, the mapping matrix from logistics supplier features to the unified semantic space is 9×6. Hierarchical contrastive learning loss is minimized using gradient descent, and these two mapping matrices are iteratively optimized. After 500 rounds of training, the loss function value dropped from an initial value of 3.62 to 0.86, and the mapping matrix stabilized.

[0047] Finally, the multi-modal feature representations are projected to a unified semantic space using the optimized semantic mapping matrix to obtain the fused feature representations. Taking manufacturing enterprise A as an example, the 9-dimensional multi-modal feature representation of enterprise A is multiplied by the 9x6 mapping matrix to obtain a 6-dimensional fused feature representation [2.31, 1.95, 2.08, 1.83, 2.15, 2.02]; similarly, the 9-dimensional multi-modal feature representation of logistics supplier B is multiplied by the corresponding 9x6 mapping matrix to obtain a 6-dimensional fused feature representation [2.18, 2.03, 1.96, 1.89, 2.24, 1.97]. These fused feature representations are located in the same semantic space, which facilitates direct similarity calculation and subsequent matching operations.

[0048] In practical applications, the trained semantic mapping matrix can be stored for rapid mapping of new enterprise and supplier features. For example, when a new manufacturing enterprise D joins the system, after obtaining its multi-modal feature representation, the mapping matrix is directly projected to the unified semantic space to obtain the fused feature representation, and then the similarity is calculated with the fused feature representations of existing logistics suppliers to quickly find potential matching suppliers.

[0049] In this paper, the feature correlation evaluation function and the hierarchical contrastive learning loss function are constructed to effectively solve the problem of multi-modal feature cross-domain fusion. The introduction of dynamic difficulty weight makes the model pay more attention to sample pairs that are difficult to distinguish, improving the learning efficiency. The hierarchical design fully utilizes the unique semantic information of features at different levels, and the differential setting of the temperature parameter adapts to the distribution characteristics of features at different levels. The semantic mapping matrix eliminates the semantic gap between modalities, realizing the deep fusion of multi-modal features. This method not only improves the discriminability of feature representations, but also preserves the semantic information of original features, providing a more accurate and comprehensive feature basis for subsequent matching recommendation.

[0050] Optionally, The step of constructing the matching degree score between the manufacturing enterprise and the logistics supplier based on the state vector includes: The historical cooperation record is obtained through a time sequence attention mechanism, and the time sequence attention mechanism calculates time sequence attention weights based on the correlation between the historical cooperation record and the time sequence context feature; Construct a multidimensional reward function, which includes a service matching reward, a cost-effectiveness reward, and a stability reward; the service matching reward is calculated based on the fusion feature similarity and historical service quality scores, the cost-effectiveness reward includes the deviation between the predicted cost and the expected cost, and the stability reward is calculated based on the cooperation volatility and the historical cooperation duration; calculate the immediate reward value based on the multidimensional reward function, and calculate the long-term reward value through temporal discount accumulation; Constructing a dual value network architecture of a state value network and an advantage function network, and calculating the Q value based on the state value output by the state value network and the action advantage output by the advantage function network; Constructing a multi-objective optimization function, the multi-objective optimization function including value loss, policy loss, and entropy loss, wherein the value loss is calculated based on the deviation between the long-term reward value and the state value, the policy loss is calculated based on the action probability distribution and the advantage function, and the entropy loss is used to adjust policy exploration; optimizing the network parameters of the dual-value network architecture based on the multi-objective optimization function; The matching score between the enterprise and the logistics supplier is calculated based on the state value and Q value output by the optimized state value network.

[0051] For example, historical collaboration records are obtained through a temporal attention mechanism. For any pair of manufacturers and logistics suppliers, the time series data of their historical collaborations is collected, including information such as collaboration frequency, order amount, and service rating. For example, the collaboration records of an electronic equipment manufacturer A and logistics supplier B over the past 12 months are as follows: the number of collaborations is [3, 2, 4, 3, 5, 4, 3, 5, 6, 4, 5, 6], the average monthly order amount (in 10,000 yuan) is [15, 12, 18, 16, 22, 20, 17, 23, 25, 19, 24, 28], and the service rating (out of 5 points) is [4.2, 4.0, 4.3, 4.1, 4.5, 4.4, 4.2, 4.6, 4.7, 4.3, 4.5, 4.8]. The temporal attention mechanism calculates the temporal attention weight based on the correlation between historical collaboration records and temporal context features. Temporal context features include seasonal factors, industry prosperity, and enterprise operating conditions. For example, the electronics industry usually has a large order volume in the fourth quarter, so the temporal context feature value of the fourth quarter is higher and is set to 0.8; while the order volume in the first quarter is usually smaller, so its temporal context feature value is lower and is set to 0.4.

[0052] Temporal attention weights are calculated by calculating the dot product of historical collaboration records and temporal context features, then normalized using a softmax function. For the 12-month historical data between Enterprise A and Supplier B, the calculated temporal attention weights are [0.03, 0.02, 0.06, 0.05, 0.09, 0.08, 0.07, 0.12, 0.15, 0.08, 0.10, 0.15] Based on these weights, the historical collaboration records are weighted and summed to obtain a representation of historical collaboration that takes temporal importance into account. For example, the weighted average monthly number of collaborations is 435, the average monthly order value is 214,200 yuan, and the average service rating is 4.47.

[0053] The fused feature representation is concatenated with the processed historical collaboration records to construct a state vector. For example, for company A and supplier B, their fused feature representations are [2.31, 1.95, 2.08, 1.83, 2.15, 2.02] and [2.18, 2.03, 1.96, 1.89, 2.24, 1.97], respectively. Their historical collaboration records are represented as [4.35, 21.42, 4.47]. Therefore, the state vector is [2.31, 1.95, 2.08, 1.83, 2.15, 2.02, 2.18, 2.03, 1.96, 1.89, 2.24, 1.97, 4.35, 21.42, 4.47].

[0054] A multi-dimensional reward function is constructed, including service matching rewards, cost-effectiveness rewards, and stability rewards. The service matching reward is calculated based on the fused feature similarity and the historical service quality score. The fused feature similarity is calculated using cosine similarity. For example, the fused feature similarity between enterprise A and supplier B is 0.92; the historical service quality score is 4.47 out of 5, normalized to 0.894. The service matching reward is calculated as the weighted sum of the fused feature similarity and the normalized service quality score, with weights of 0.6 and 0.4, respectively, resulting in 0.92 × 0.6 + 0.894 × 0.4 = 0.91. The cost-effectiveness reward includes the deviation between the predicted cost and the expected cost. Forecasted costs are estimated using historical order data and current market prices. For example, if Company A's forecasted logistics costs for Supplier B are 5,000 yuan per order, and the expected costs are determined based on the company's budget and industry standards, for example, if Company A's expected logistics costs are 4,800 yuan per order, then the cost deviation rate is (5,000 - 4,800) / 4,800 = 0.042. The cost-effectiveness bonus also considers a comprehensive benefit index, comprising historical logistics efficiency improvements and cost savings. The efficiency improvement rate is calculated as the percentage reduction in logistics cycle time, for example, 8%, while the cost savings rate is calculated as the percentage reduction in logistics costs, for example, 5%. The comprehensive benefit index is weighted to 0.13. The cost-effectiveness bonus also includes a market environment factor, the regional logistics price index, which is 0.96. A comprehensive evaluation of these three factors yields a final cost-effectiveness bonus value of 0.125, calculated as (1 - 0.042) × 0.13 × 0.96 = 0.125.

[0055] Stability bonuses are calculated based on the volatility of collaboration and the historical duration of collaboration. Collaboration volatility is calculated by dividing the standard deviation of the historical collaboration frequency by the average. For example, if the standard deviation of the collaboration frequency between Company A and Supplier B is 1.23 and the average is 4.35, the volatility is 1.23 / 4.35 = 0.283. The historical duration of collaboration is 12 months, normalized to 0.6 (assuming the longest collaboration period is 20 months). The stability bonus is calculated as 1 minus the volatility multiplied by the normalized collaboration duration, which equals (1-0.283) × 0.6 = 0.43.

[0056] The immediate reward value is the weighted sum of the three-dimensional rewards, with weights of 0.5, 0.3, and 0.2, respectively, resulting in 0.91×0.5+0.125×0.3+0.43×0.2=0.588. The long-term reward value is calculated by accumulating the temporal discount factor, with the temporal discount factor set to 0.9. Assuming that the immediate reward predictions for the next three time steps are 0.595, 0.603, and 0.612, respectively, the long-term reward value is 0.588+0.9×0.595+0.9 2 ×0.603+0.9 3 ×0.612=2.185.

[0057] The input of the state value network is the state vector, and the output is the state value V(s). The state vector consists of a fusion feature representation and a historical cooperation record, and its dimension is 15. Taking an electronic equipment manufacturer A and a logistics supplier B as an example, their state vectors are [2.31, 1.95, 2.08, 1.83, 2.15, 2.02, 2.18, 2.03, 1.96, 1.89, 2.24, 1.97, 4.35, 21.42, 4.47]. The state value network adopts a multi-layer perceptron structure, which includes three hidden layers. The number of neurons in each layer is 64, 32, and 16 respectively. The activation functions are ReLU, ReLU, and tanh respectively. The final output layer is a single neuron, using a linear activation function. The advantage function network structure is similar to the state value network. It also inputs a state vector, but the output is the advantage value A(s, a) of each action. In this embodiment, the action space contains two discrete actions: matching and mismatching. The network consists of three hidden layers, with the same number of neurons as the state-value network. The output layer has two neurons, corresponding to two actions. The network is initialized using the Xavier method and optimized using the Adam algorithm with a learning rate of 0.001. To prevent overfitting, a dropout layer with a dropout rate of 0.2 is added after each hidden layer. Taking Company A and Supplier B as an example, the state value V(s) calculated through forward propagation is 2.05; the matching action advantage is 0.35, and the mismatching action advantage is -0.28. The Q value is calculated based on the state value and action advantage: first, the average action advantage is calculated as (0.35 + (-0.28)) / 2 = 0.035. Then, the Q value of the matching action is 2.05 + 0.35 - 0.035 = 2.365; the Q value of the mismatching action is 2.05 + (-0.28) - 0.035 = 1.735.

[0058] Construct a multi-objective optimization function, including value loss, policy loss, and entropy loss. The value loss is calculated based on the deviation between the long-term reward value and the state value, for example (2.185-2.05) 2 = 0.018; the policy loss is calculated based on the action probability distribution and the advantage function. For example, if the probability of matching an action is 0.85 and the advantage value is 0.35, the policy loss is -0.85 × ln(0.85) × 0.35 = 0.048; the entropy loss is used to adjust the policy exploration, for example, -[0.85 × ln(0.85) + 0.15 × ln(0.15)] × 0.1 = 0.042. The multi-objective optimization function is the weighted sum of the three losses, with weights of 0.5, 0.3, and 0.2, respectively, resulting in 0.018 × 0.5 + 0.048 × 0.3 + 0.042 × 0.2 = 0.031.

[0059] Based on the multi-objective optimization function, the network parameters of the dual value network architecture are optimized using the stochastic gradient descent method. The learning rate is set to 0.001, the batch size is 64, and the training rounds are 200 rounds. During the optimization process, the multi-objective loss decreases from the initial 0.285 to 0.031, indicating that the network parameters gradually tend to be optimal.

[0060] Finally, the matching score between the enterprise and the logistics supplier is calculated based on the state value and Q value output by the optimized state value network. The matching score is calculated by weighted combination of state value and matching action Q value, with weights of 0.4 and 0.6 respectively, for example, 2.05 x 0.4 + 2.365 x 0.6 = 2.239, and then normalized to the interval of 0-100, to get the final matching score of 88.7 points.

[0061] For all pairs of manufacturing enterprises and logistics suppliers, the matching score is calculated using the above method. For example, for enterprise A and supplier C, the matching score is 65.3, and for supplier D, the matching score is 72.8, and for supplier B, the matching score is 88.7, which is the highest, indicating that supplier B is the most suitable logistics service provider for enterprise A.

[0062] In this paper, through the combination of time series attention mechanism, multi-dimensional reward function and dual value network architecture, the precise evaluation of the matching relationship between manufacturing enterprises and logistics suppliers is realized. The time series attention mechanism effectively captures the time series patterns and importance differences in historical cooperation data, the multi-dimensional reward function comprehensively considers key factors such as service matching degree, cost benefit and stability, and the dual value network architecture provides more stable and accurate value evaluation. Balancing short-term interests and long-term development needs, a more scientific and reasonable matching evaluation system is built for manufacturing enterprises and logistics suppliers.

[0063] Optionally, The steps of constructing the multi-dimensional reward function include: Based on the similarity of the fusion features, the basic service matching degree is obtained, and combined with the historical service quality influence factor based on the time series decay weight and the adaptive scale factor based on the current business scale, the service matching degree reward is obtained; Based on the deviation of the predicted cost and the expected cost, the cost deviation evaluation value is calculated, and combined with the comprehensive benefit index based on the historical logistics efficiency improvement rate and the cost saving rate, and the market environment factor based on the regional logistics price index, the cost benefit reward is obtained; Based on the historical cooperation index, the cooperation fluctuation degree is calculated, and combined with the influence factor based on the cooperation duration and the seasonal adjustment factor based on the cycle fluctuation, the stability reward is obtained; The weight coefficients of rewards in each dimension are dynamically adjusted through gradient updates. The rewards in the three dimensions are weighted and combined based on the dynamic weights to obtain the immediate reward value. A time domain adjustment coefficient is introduced to obtain the long-term reward value through temporal discount accumulation.

[0064] For example, the basic service matching degree is obtained based on the fused feature similarity. This fused feature similarity is calculated by calculating the cosine similarity between the fused feature representations of the manufacturer and the logistics provider. For example, for electronic equipment manufacturer A and logistics provider B, their fused feature representations are [2.31, 1.95, 2.08, 1.83, 2.15, 2.02] and [2.18, 2.03, 1.96, 1.89, 2.24, 1.97], respectively. The calculated cosine similarity is 0.92, which is used as the basic service matching degree.

[0065] The time-series decay weight is explicitly calculated using an exponential decay function: the weight at the current time point is 1.0, and the weight at the previous i-th time point is power(decay_rate, i), where decay_rate is the decay rate, which is set to 0.85. For example, the service ratings (out of 5) of Enterprise A and Supplier B in the past 6 months are [4.2, 4.5, 4.3, 4.6, 4.7, 4.8], and the corresponding time-series decay weights are [0.85 5 =0.44,0.85 4 =0.52,0.85 3 =0.61,0.85 2 =0.72,0.85 1 =0.85, 1.0]. The weighted average historical service score is (4.2×0.44+4.5×0.52+4.3×0.61+4.6×0.72+4.7×0.85+4.8×1.0) / (0.44+0.52+0.61+0.72+0.85+1.0)=4.57. The historical service quality impact factor is the normalized score value, that is, 4.57 / 5=0.914.

[0066] The adaptive scale factor is calculated quantitatively using the Sigmoid function: 1 / (1+exp(-k×(current_scale / average_scale-1))), where "current_scale" represents the company's current business scale indicator, such as the current average monthly logistics order volume; "average_scale" represents the historical average business scale indicator, such as the historical average monthly logistics order volume; and k is a sensitivity parameter set to 2. Company A's current average monthly logistics order volume is 80, while its historical average order volume is 60. The ratio is 80 / 60 = 1.33. Substituting this into the formula, the adaptive scale factor is 1 / (1+exp(-2×(1.33-1))) = 1 / (1+exp(-0.66)) = 1 / (1+0.517) = 0.659.

[0067] The service matching reward is calculated using the weighted fusion formula: (basic service matching × w1 + historical service quality impact factor × w2) × (1 + adaptive scale factor × α), where w1 = 0.6, w2 = 0.4, and α = 0.3. Substituting these values, the service matching reward is (0.92 × 0.6 + 0.914 × 0.4) × (1 + 0.659 × 0.3) = 0.918 × 1.198 = 1.099.

[0068] Forecasted costs are estimated using historical order data and current market prices. For example, if Company A predicts a logistics cost of 5,000 yuan per order from Supplier B, and the expected cost is determined based on the company's budget and industry standards, for example, if Company A expects a logistics cost of 4,800 yuan per order, the cost deviation rate is (5,000 yuan - 4,800 yuan) / 4,800 yuan = 0.042. A comprehensive benefit index is constructed based on historical logistics efficiency improvement and cost savings. The efficiency improvement rate is calculated by the percentage reduction in logistics cycle time, and the cost savings rate is calculated by the percentage reduction in logistics costs. The comprehensive benefit index uses a weighted average method, with efficiency improvement accounting for 60% and cost savings accounting for 40%. For example, if the average logistics cycle time has shortened by 8% and logistics costs have decreased by 5% over the past six months, the comprehensive benefit index is 0.068. A market environment factor is constructed based on the regional logistics price index. The regional logistics price index is calculated by comparing current regional logistics prices with baseline prices. The market environment factor is determined based on the change in the current price index relative to the baseline value. For example, if the current regional logistics price index is 105 (relative to a benchmark value of 100), the market environment factor is 0.95, indicating that a slight price increase has led to a slightly unfavorable market environment. Based on the cost deviation rate, the cost deviation assessment value can be calculated as 1-0.042 = 0.958. The final cost-effectiveness bonus is calculated by weighting the three factors: the cost deviation assessment value is assigned a 50% weight, the comprehensive benefit index is assigned a 30% weight, and the market environment factor is assigned a 20% weight. The calculated cost-effectiveness bonus is 0.958 × 0.5 + 0.068 × 0.3 + 0.95 × 0.2 = 0.479 + 0.0204 + 0.19 = 0.6894.

[0069] Construct a stability reward. The volatility of cooperation is quantified by the coefficient of variation: the weighted average of the coefficients of variation of each historical cooperation indicator. The standard deviation of the number of cooperations between enterprise A and supplier B in the past 12 months is 1.23, the average is 4.17, and the coefficient of variation is 1.23 / 4.17=0.295; the coefficient of variation of order amount is 0.232; and the coefficient of variation of service score is 0.065. The indicator weights are 0.3, 0.3, and 0.4 respectively, and the volatility of cooperation is 0.295×0.3+0.232×0.3+0.065×0.4=0.184. The impact factor of cooperation duration is calculated by logarithmic normalization. For example, enterprise A and supplier B have been cooperating for 12 months, and the industry reference longest cooperation period is 60 months. Substituting into the formula, the impact factor of cooperation duration is ln(12) / ln(60)=2.485 / 4.094=0.607. The seasonal adjustment factor is quantified using Fourier analysis: 1 - seasonal_index × adjustment_coefficient, where seasonal_index is the seasonal fluctuation index, obtained by extracting the seasonal component through Fourier transform of historical data; adjustment_coefficient is the adjustment coefficient, set to 0.5. Company A's seasonal fluctuation index is 0.28. Substituting this into the formula yields a seasonal adjustment factor of 1 - 0.28 × 0.5 = 0.86. The stability bonus is calculated using the formula: (1 - collaboration volatility) × collaboration duration impact factor × seasonal adjustment factor. Substituting this value into the formula yields a stability bonus of (1 - 0.184) × 0.607 × 0.86 = 0.425.

[0070] The weighting coefficients for each dimension of rewards are determined through iterative optimization. Initial weights are set, and then the direction and magnitude of adjustments are calculated based on the difference between historical matching results and predicted results. Each adjustment is made by incrementally adjusting the weights in small steps of 0.01 to ensure stable system convergence. After multiple rounds of iterative optimization, based on extensive historical matching data, the final weightings for service matching rewards are 0.45, cost-effectiveness rewards are 0.35, and stability rewards are 0.2.

[0071] The immediate reward value is calculated by linear weighting: service matching reward × w1 + cost-effectiveness reward × w2 + stability reward × w3. Substituting the numerical values, the immediate reward value is 1.099 × 0.45 + 0.072 × 0.35 + 0.425 × 0.2 = 0.579.

[0072] The time adjustment factor is set based on the time period: 1.2 for peak season, 1.0 for standard season, and 0.8 for off-season. Currently in the fourth quarter (peak season), the time adjustment factor is 1.2, and the adjusted instant reward value is 0.579 × 1.2 = 0.695.

[0073] The long-term reward value is calculated by discount accumulation: immediate reward + γ × next moment reward + γ 2 ×The next two moments reward +... +γ n × reward at the nth moment, where γ is the discount factor, set to 0.9. The current reward is 0.695, and the predicted rewards for the next three time steps are 0.705, 0.712, and 0.720 respectively. Substituting into the formula, the long-term reward value is 0.695+0.9×0.705+0.9 2 ×0.712+0.9 3 ×0.720=2.517.

[0074] In practice, each parameter can be adjusted based on the characteristics of different industries and the needs of enterprises. For example, industries with high requirements for logistics timeliness can increase the weight of service matching rewards; cost-sensitive enterprises can increase the weight of cost-effectiveness rewards.

[0075] This technical solution comprehensively evaluates the matching quality between manufacturing companies and logistics providers by constructing a multi-dimensional reward function. Each calculation factor is quantified using a clear mathematical formula, including time-decrease weights, adaptive scale factors, expected returns, market environment adaptation factors, and seasonal adjustment factors. A dynamic weight adjustment mechanism, combined with temporal adjustments and temporal discounts, effectively balances short-term benefits with long-term value, providing a precise optimization target for deep reinforcement learning algorithms. This significantly improves the accuracy, stability, and adaptability of intelligent matching in manufacturing logistics, enabling adaptive adjustments to matching strategies based on complex situations such as changes in business scale, market fluctuations, and seasonal demand differences.

[0076] Optional, The steps of dividing the logistics orders into time windows and optimizing the allocation based on the matching scores to obtain the optimized allocation results include: Obtain a set of orders to be assigned and calculate the order density within each time period. The order density includes the number of orders per unit time, the uniformity of order distribution, and the order resource occupancy rate. A piecewise function is used to set the time window benchmark size based on the order density. Calculating the urgency of each order based on the order delivery deadline, order priority, and associated order constraints; adaptively adjusting the time window benchmark size based on the urgency, where a high urgency corresponds to a time window contraction factor and a low urgency corresponds to a time window expansion factor; Within the adjusted time window, order allocation constraints are set based on the matching score, including time window constraints, capacity constraints, and order association constraints; the integer programming method is used to optimize the allocation of logistics orders to obtain the optimized allocation results.

[0077] For example, a set of orders to be assigned is obtained, and the order density in each time period is calculated. Order density includes three indicators: the number of orders per unit time, the uniformity of order distribution, and the order resource occupancy rate. The number of orders per unit time directly counts the total number of orders in a specific time period (such as every hour). The uniformity of order distribution is determined by calculating the degree of dispersion of the order distribution within the time period. The smaller the value, the more uniform the distribution. The order resource occupancy rate calculates the ratio of the total amount of transportation resources occupied by each order to the total amount of available resources. For example, during the morning rush hour of 8:00-9:00, the number of orders in a certain area is 120, the order distribution uniformity is 0.15 (indicating relatively uniform), and the order resource occupancy rate is 0.75 (indicating high resource utilization).

[0078] A piecewise function is used to set the time window benchmark size based on order density. When the order density is higher than the threshold, a smaller time window is set to improve scheduling accuracy; when the order density is lower than the threshold, a larger time window is set to improve resource utilization. For example, when the number of orders is greater than 100 orders / hour and the resource occupancy rate is greater than 0.7, the benchmark time window is set to 30 minutes; when the number of orders is between 50-100 orders / hour or the resource occupancy rate is between 0.4-0.7, the benchmark time window is set to 60 minutes; when the number of orders is less than 50 orders / hour and the resource occupancy rate is less than 0.4, the benchmark time window is set to 120 minutes. For example, for the morning rush hour mentioned above, since the number of orders is 120 orders / hour and the resource occupancy rate is 0.75, the benchmark time window is set to 30 minutes according to the piecewise function.

[0079] The urgency of each order is calculated based on the order's delivery deadline, order priority, and associated order constraints. The order delivery deadline is determined by the difference between the order's promised delivery deadline and the current time; a smaller difference indicates greater urgency. Order priority is set based on order type (such as VIP orders, promotional orders, or regular orders), with higher-priority orders receiving a higher urgency. Associated order constraints consider dependencies between orders, such as orders in the same batch or requiring centralized delivery. Urgency is quantified as a value between 0 and 1, with higher values ​​indicating greater urgency. For example, Order A has two hours until its delivery deadline, is a VIP order (priority coefficient of 1.2), and has no associated order constraints. Its urgency is calculated as 0.85. Order B has five hours until its delivery deadline, is a regular order (priority coefficient of 1.0), and has associated orders (association coefficient of 1.1). Its urgency is calculated as 0.42.

[0080] Adaptively adjust the baseline time window size based on urgency. Set an urgency threshold. When the urgency is above the upper threshold (e.g., 0.7), the time window is reduced by a contraction factor (e.g., 0.6). When the urgency is below the lower threshold (e.g., 0.3), the time window is expanded by an expansion factor (e.g., 1.5). When the urgency is between the two thresholds, the time window remains at the baseline size. For example, for order A with an urgency of 0.85, its time window should be reduced from the baseline 30 minutes to 30 × 0.6 = 18 minutes. For order B with an urgency of 0.42, its time window remains at the baseline 60 minutes.

[0081] Within the adjusted time window, order allocation constraints are set based on the matching score. Time window constraints ensure that orders must be allocated within the specified time window. Capacity constraints ensure that orders allocated within each time window do not exceed the maximum available capacity. Order association constraints consider the dependencies between orders and mainly fall into three categories: same-batch orders, same-destination orders, and time-dependent orders. Same-batch orders refer to multiple orders from the same customer and require centralized delivery to improve the customer experience. Same-destination orders refer to orders destined for the same region and should be centrally allocated to improve delivery efficiency. Time-dependent orders refer to orders with a sequential order, such as first-come, first-served orders. A weight is assigned to each constraint; for example, a time window constraint weight of 0.5, a capacity constraint weight of 0.3, and an order association constraint weight of 0.2. Based on the matching score, a supplier with a matching score above 0.8 is prioritized; a supplier with a matching score below 0.5 is excluded from the candidate list.

[0082] The integer programming approach models the order assignment problem as an integer programming problem. The decision variable x_ij indicates whether order i is assigned to supplier j and takes the value 0 or 1. The objective function is a weighted sum of maximizing the overall matching score and minimizing the delivery cost, expressing a balance between efficiency and cost. In the objective function, the matching component is calculated as the sum of the matching scores of all assigned orders, and the delivery cost component is calculated as the sum of the estimated delivery costs of all assigned orders. These two components are weighted by weight coefficients. Typically, the matching weight is set to α = 0.7 and the delivery cost weight is set to β = 0.3. The weights can be adjusted based on business needs. Constraints include: each order must be assigned to exactly one supplier; the number of orders per supplier within a specific time window must not exceed its maximum processing capacity; strongly correlated orders must be assigned to the same supplier; and orders must be assigned within the specified time window. The solution uses a branch-and-bound algorithm using an integer programming solver (such as CPLEX or Gurobi).

[0083] Based on the optimized allocation results obtained through integer programming, the results are processed and logistics suppliers are recommended to the manufacturing enterprise. Result processing consists of three steps: allocation result verification, result optimization, and result conversion. Allocation result verification checks whether all constraints are satisfied, specifically whether strongly correlated orders are assigned to the same supplier and whether time window constraints are met. If any constraints are found to be unsatisfied, a local adjustment process is triggered. For example, if orders A and B, though strongly correlated, are assigned to different suppliers, the system reassigns them to the supplier with the higher matching score. Result optimization fine-tunes the results for specific business scenarios. For example, for frequently collaborated suppliers, the number of allocated orders can be appropriately increased to maintain long-term partnerships; for newly introduced suppliers, a small number of orders can be allocated to evaluate their capabilities. In specific implementations, the adjustment ratio is set to no more than 10% of the total order volume to ensure that overall optimality is not significantly affected. Result conversion transforms the optimized allocation results into actionable recommendations. For each manufacturing enterprise, a list of recommended suppliers is generated, consisting of a primary recommended supplier and alternative suppliers. The primary recommended supplier is the supplier assigned to the order according to the integer programming results, while the alternative suppliers are the second to fourth most compatible suppliers. For example, the main recommended supplier for order C is supplier X (matching degree 0.92), and the alternative suppliers include supplier Y (matching degree 0.85) and supplier Z (matching degree 0.79).

[0084] This paper implements dynamic time window partitioning based on order density and adaptive adjustment based on urgency, combining it with matching scores to optimize order allocation. This effectively addresses the issues of fixed time windows and uneven resource allocation in traditional logistics order allocation. By dynamically adjusting the time window size, it not only meets the timely delivery needs of high-urgency orders, but also improves resource utilization during low-density periods. At the same time, the constraints based on matching scores ensure the rationality and efficiency of order allocation, ultimately achieving the dual goals of improving logistics efficiency and reducing costs.

[0085] Secondly, we provide a manufacturing logistics intelligent matching recommendation system based on multimodal data, including: The first unit is used to obtain logistics demand data of manufacturing enterprises and service capability data of logistics suppliers; the second unit is used to construct a multi-level feature attention network to extract features from the logistics demand data and the service capability data, and the multi-level feature attention network includes an enterprise scale layer, an industry attribute layer, and a business feature layer; the multi-level feature attention network is trained based on historical matching data to obtain feature importance scores, and features are screened according to the feature importance scores to obtain multimodal feature representations; the third unit is used to construct positive and negative sample pairs based on the multimodal feature representations, and obtain an inter-modal semantic mapping matrix through hierarchical contrast learning; the multimodal feature representations are projected into a unified semantic space using the semantic mapping matrix to obtain a fused feature representation; the fourth unit is used to construct the fused feature representation and historical cooperation records into a state vector; based on the state vector, an immediate reward function and a long-term reward function are constructed, and a matching score between the manufacturing enterprise and the logistics supplier is obtained by optimizing the deep reinforcement learning algorithm; the fifth unit is used to divide the logistics orders into time windows and optimize the allocation based on the matching scores to obtain an optimized allocation result; based on the optimized allocation result, a logistics supplier is recommended to the manufacturing enterprise. According to a third aspect, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

Claims

1. The intelligent matching recommendation method for manufacturing logistics based on multimodal data is characterized by: include: Obtain logistics demand data of manufacturing enterprises and service capability data of logistics suppliers; Constructing a multi-level feature attention network to extract features from the logistics demand data and the service capability data, wherein the multi-level feature attention network includes an enterprise scale layer, an industry attribute layer, and a business feature layer; training the multi-level feature attention network based on historical matching data to obtain feature importance scores, and screening features based on the feature importance scores to obtain a multimodal feature representation; constructing positive and negative sample pairs based on the multimodal feature representation, and obtaining an inter-modal semantic mapping matrix through a hierarchical contrastive learning approach; and utilizing the semantic mapping matrix to project the multimodal feature representation into a unified semantic space to obtain a fused feature representation; The fused feature representation and historical cooperation records are constructed into a state vector; an immediate reward function and a long-term reward function are constructed based on the state vector, and a matching score between the manufacturing enterprise and the logistics supplier is obtained through optimization using a deep reinforcement learning algorithm; Divide the logistics orders into time windows and optimize their allocation based on the matching scores to obtain optimized allocation results; Based on the optimized allocation results, a logistics supplier is recommended to the manufacturing enterprise.

2. The method according to claim 1, characterized in that Constructing a multi-level feature attention network to extract features from the logistics demand data and the service capability data, wherein the multi-level feature attention network includes an enterprise scale layer, an industry attribute layer, and a business feature layer; The steps of training the multi-level feature attention network based on historical matching data to obtain feature importance scores, screening features according to the feature importance scores, and obtaining a multimodal feature representation include: The logistics demand data and the service capability data are constructed into enterprise scale layer features, industry attribute layer features, and business feature layer features according to time series characteristics; in the multi-level feature attention network, a scale layer attention calculation unit is constructed based on the enterprise scale layer features, and a feature stability matrix is ​​introduced to calculate the enterprise scale layer feature weights; an industry layer attention calculation unit is constructed based on the industry attribute layer features, and an industry correlation strength matrix is ​​introduced to calculate the industry attribute layer feature weights; a business layer attention calculation unit is constructed based on the business feature layer features, and a time decay factor is introduced to calculate the business feature layer feature weights; Calculating inter-layer downward enhancement features and inter-layer upward feedback features based on the enterprise scale layer feature weight, the industry attribute layer feature weight, and the business feature layer feature weight; performing weighted fusion on the inter-layer downward enhancement features and the inter-layer upward feedback features to obtain inter-layer interaction features; A multi-objective loss function is constructed based on historical matching data to train the multi-level feature attention network, wherein the multi-objective loss function includes a matching accuracy loss term, a feature stability loss term, and a temporal consistency loss term; the feature independent importance score and the feature interaction importance score are calculated, and the feature importance score is obtained based on the feature independent importance score and the feature interaction importance score; the inter-layer interaction features are screened according to the feature importance score to obtain a multimodal feature representation.

3. The method according to claim 2, characterized in that Calculate inter-layer downward enhancement features and inter-layer upward feedback features based on the enterprise scale layer feature weight, the industry attribute layer feature weight, and the business feature layer feature weight; The step of weightedly fusing the inter-layer downward enhancement feature and the inter-layer upward feedback feature to obtain the inter-layer interaction feature includes: Multiplying the matrices of the enterprise scale layer feature weights, the industry attribute layer feature weights, and the business feature layer feature weights to obtain an inter-layer interaction matrix; calculating the complementarity scores between the feature representations corresponding to the enterprise scale layer feature weights, the industry attribute layer feature weights, and the business feature layer feature weights, and calculating the inter-layer interaction strength based on the complementarity scores and the inter-layer interaction matrix; Constructing an information flow gating unit, wherein the information flow gating unit calculates an information flow coefficient based on the characteristic hidden states of adjacent levels; The inter-layer interaction intensity, the information flow coefficient, the enterprise scale layer feature weight, the industry attribute layer feature weight, and the business feature layer feature weight are combined to calculate the inter-layer downward enhancement feature and the inter-layer upward feedback feature, wherein the inter-layer downward enhancement feature is obtained by the enhancement effect of the high-level feature on the low-level feature, and the inter-layer upward feedback feature is obtained by the feedback effect of the low-level feature on the high-level feature; A feature fusion function is constructed, a dynamic fusion weight is calculated by the feature fusion function, and inter-layer downward enhancement features and inter-layer upward feedback features are weightedly fused according to the dynamic fusion weight to obtain inter-layer interaction features.

4. The method according to claim 1, wherein The steps of constructing positive and negative sample pairs based on the multimodal feature representation, obtaining an inter-modal semantic mapping matrix through a hierarchical contrastive learning method, and projecting the multimodal feature representation into a unified semantic space using the semantic mapping matrix to obtain a fused feature representation include: Constructing a positive sample pair set and a negative sample pair set based on the multimodal feature representation, wherein the positive sample pair set includes matching demand-side feature and supply-side feature pairs, and the negative sample pair set includes mismatched demand-side feature and supply-side feature pairs; Constructing a feature relevance evaluation function that simultaneously considers feature space similarity and industry collaboration frequency; and calculating a dynamic difficulty weight for a sample pair based on the feature relevance evaluation function. Constructing a hierarchical contrastive learning loss function, wherein the hierarchical contrastive learning loss function includes multiple feature levels, each feature level is assigned a corresponding similarity metric function and a temperature parameter; weighting the positive and negative sample pairs of each feature level according to the dynamic difficulty weight to obtain a weighted hierarchical contrastive learning loss; and dynamically adjusting the weight coefficients of different feature levels based on gradient updates; The semantic mapping matrix is ​​optimized based on the hierarchical contrastive learning loss function, and the multimodal feature representation is projected into a unified semantic space through the semantic mapping matrix to obtain a fused feature representation.

5. The method according to claim 1, wherein The steps of constructing the fused feature representation and historical cooperation records into a state vector; constructing an immediate reward function and a long-term reward function based on the state vector, and optimizing the deep reinforcement learning algorithm to obtain a matching score between the manufacturing enterprise and the logistics supplier include: The historical collaboration records are obtained through a temporal attention mechanism, and the temporal attention mechanism calculates a temporal attention weight based on the correlation between the historical collaboration records and temporal context features; Construct a multidimensional reward function, which includes a service matching reward, a cost-effectiveness reward, and a stability reward; the service matching reward is calculated based on the fusion feature similarity and historical service quality scores, the cost-effectiveness reward includes the deviation between the predicted cost and the expected cost, and the stability reward is calculated based on the cooperation volatility and the historical cooperation duration; calculate the immediate reward value based on the multidimensional reward function, and calculate the long-term reward value through temporal discount accumulation; Constructing a dual value network architecture of a state value network and an advantage function network, and calculating the Q value based on the state value output by the state value network and the action advantage output by the advantage function network; Constructing a multi-objective optimization function, the multi-objective optimization function including value loss, policy loss, and entropy loss, wherein the value loss is calculated based on the deviation between the long-term reward value and the state value, the policy loss is calculated based on the action probability distribution and the advantage function, and the entropy loss is used to adjust policy exploration; optimizing the network parameters of the dual-value network architecture based on the multi-objective optimization function; The matching score between the enterprise and the logistics supplier is calculated based on the state value and Q value output by the optimized state value network.

6. The method according to claim 5, characterized in that The steps to construct a multi-dimensional reward function include: The basic service matching degree is obtained based on the fusion feature similarity, and the service matching degree reward is obtained by combining the historical service quality impact factor based on the time-series decay weight and the adaptive scale factor based on the current business scale; The cost deviation assessment value is calculated based on the deviation between the predicted cost and the expected cost, and combined with the comprehensive benefit index based on the historical logistics efficiency improvement rate and cost saving rate and the market environment factor based on the regional logistics price index to obtain the cost-effectiveness reward; The volatility of cooperation is calculated based on historical cooperation indicators, and the impact factor based on the duration of cooperation and the seasonal adjustment factor based on cyclical fluctuations are combined to obtain stability rewards; The weight coefficients of rewards in each dimension are dynamically adjusted through gradient updates. The rewards in the three dimensions are weighted and combined based on the dynamic weights to obtain the immediate reward value. A time domain adjustment coefficient is introduced to obtain the long-term reward value through temporal discount accumulation.

7. The method according to claim 1, characterized in that The steps of dividing the logistics orders into time windows and optimizing the allocation based on the matching scores to obtain the optimized allocation results include: Obtain a set of orders to be assigned and calculate the order density within each time period. The order density includes the number of orders per unit time, the uniformity of order distribution, and the order resource occupancy rate. A piecewise function is used to set the time window benchmark size based on the order density. Calculating the urgency of each order, the urgency being calculated based on the order delivery deadline, order priority, and associated order constraints; and adaptively adjusting the time window benchmark size based on the urgency; Within the adjusted time window, order allocation constraints are set based on the matching score, including time window constraints, capacity constraints, and order association constraints; the integer programming method is used to optimize the allocation of logistics orders to obtain the optimized allocation results.

8. A manufacturing logistics intelligent matching recommendation system based on multimodal data, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain logistics demand data of manufacturing enterprises and service capacity data of logistics suppliers; The second unit is configured to construct a multi-level feature attention network to extract features from the logistics demand data and the service capability data, wherein the multi-level feature attention network includes an enterprise scale layer, an industry attribute layer, and a business feature layer; the multi-level feature attention network is trained based on historical matching data to obtain feature importance scores, and features are screened according to the feature importance scores to obtain a multimodal feature representation; The third unit is used to construct positive and negative sample pairs based on the multimodal feature representation, obtain an inter-modal semantic mapping matrix through a hierarchical contrastive learning method, and use the semantic mapping matrix to project the multimodal feature representation into a unified semantic space to obtain a fused feature representation; A fourth unit is configured to construct a state vector from the fused feature representation and historical cooperation records; construct an immediate reward function and a long-term reward function based on the state vector, and obtain a matching score between the manufacturing enterprise and the logistics supplier through deep reinforcement learning algorithm optimization; A fifth unit is configured to divide the logistics orders into time windows and optimize the allocation based on the matching scores to obtain an optimized allocation result; Based on the optimized allocation results, a logistics supplier is recommended to the manufacturing enterprise.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images