Digital enterprise management data analysis method and system based on artificial intelligence

Abstract

The invention relates to the technical field of image-text data processing, provides a digital enterprise management data analysis method and system based on artificial intelligence, and is used for generating a visual image-text analysis report on the basis of comprehensive and multivariate supply chain data by fusing dynamic decision logic, improving the comprehensiveness, dynamism and decision guidance of image-text generation and improving the user experience. Therefore, the enterprise management system is effectively guided to dynamically adjust the supply chain. The method comprises the steps of collecting a supply chain data set of a target enterprise, and performing feature extraction on the supply chain data set to obtain a supply chain trend prediction feature set and a supply chain anomaly detection feature set; based on a preset reinforcement learning decision optimization model, performing dynamic decision path generation processing on the supply chain trend prediction feature set and the supply chain anomaly detection feature set to generate a multi-stage decision optimization path; and generating a visual image-text analysis report according to the multi-stage decision optimization path and the supply chain data set, and feeding back the visual image-text analysis report to the enterprise management system.

Description

Technical Field

[0001] The present invention belongs to the technical field of graphic and text data processing, and specifically relates to a digital enterprise management data analysis method and system based on artificial intelligence. Background Art

[0002] As enterprises advance their digital transformation, the generation of visual graphics has become increasingly crucial in the field of enterprise data analysis. Visual graphics can transform massive amounts of complex data into intuitive and easy-to-understand charts and graphs, greatly improving the efficiency of data understanding and analysis, helping business managers quickly obtain key information and make accurate decisions. Effective visual graphics can clearly illustrate every aspect of a company's operations, helping companies grasp the overall situation and optimize resource allocation.

[0003] However, current visualization technology has significant shortcomings. Most existing technologies rely on a single data source, typically relying on a single type of data for image generation. This results in a one-sided image that fails to fully reflect a company's operations. Furthermore, existing image generation lacks deep integration with complex enterprise decision-making logic, resulting in static displays. This makes it difficult to generate visualizations that are dynamic and decision-oriented, tailored to the company's real-time data and dynamic decision-making needs. Summary of the Invention

[0004] The present invention provides an artificial intelligence-based digital enterprise management data analysis method and system, which is used to generate visual graphic analysis reports based on comprehensive and diverse supply chain data and integrate dynamic decision-making logic. It can improve the comprehensiveness, dynamism and decision-making guidance of graphic generation, thereby effectively guiding the enterprise management system to dynamically adjust the supply chain.

[0005] In a first aspect, an embodiment of the present invention provides an artificial intelligence-based digital enterprise management data analysis method, which is applied to a digital enterprise management data analysis system, the method comprising: collecting a supply chain data set of a target enterprise; the supply chain data set comprises order interaction record data in text form and logistics node monitoring data in image form; performing feature extraction on the supply chain data set to obtain a supply chain trend prediction feature set and a supply chain anomaly detection feature set; based on a preset reinforcement learning decision optimization model, performing dynamic decision path generation processing on the supply chain trend prediction feature set and the supply chain anomaly detection feature set to generate a multi-stage decision optimization path; generating a visual graphic analysis report based on the multi-stage decision optimization path and the supply chain data set, and feeding the visual graphic analysis report back to the enterprise management system; wherein, the visual graphic analysis report is used to instruct the enterprise management system to perform supply chain dynamic adjustment operations.

[0006] In a second aspect, an embodiment of the present invention provides a digital enterprise management data analysis system, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0007] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, which includes a computer program. When the computer program is run on a digital enterprise management data analysis system, the computer program is used to enable the digital enterprise management data analysis system to execute the steps of the above method.

[0008] The embodiment of the present invention is based on comprehensive and diverse supply chain data and integrates dynamic decision-making logic to generate a visual graphic analysis report, which can improve the comprehensiveness, dynamism and decision-making guidance of graphic generation, thereby effectively guiding the enterprise management system to dynamically adjust the supply chain.

[0009] In detail, the embodiments of the present invention provide a comprehensive and diverse data foundation for the generation of graphics and text by collecting supply chain data sets that include text and image formats; the supply chain trend prediction feature set and supply chain anomaly detection feature set obtained through feature extraction inject depth and pertinence into the graphics and text content; the multi-stage decision optimization path generated based on the reinforcement learning decision optimization model further clarifies the logical framework for graphics and text generation, ensuring that the generated visual graphics and text analysis report is not a simple list of data, but an in-depth analysis result with logical coherence and decision-making orientation. The visual graphics and text analysis report not only intuitively presents the supply chain status, but also instructs the enterprise management system to perform dynamic supply chain adjustment operations in a clear and understandable graphic form based on the multi-stage decision path, thereby improving the visualization and intelligence level of enterprise decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A flowchart of a digital enterprise management data analysis method based on artificial intelligence provided by an embodiment of the present invention.

[0011] Figure 2 This is a structural diagram of a digital enterprise management data analysis system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the technical solutions of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments described in the present invention document without making any creative efforts shall fall within the scope of protection of the technical solutions of the present invention.

[0013] See also Figure 1 , which is an artificial intelligence-based digital enterprise management data analysis method provided in an embodiment of the present invention. The method can be applied to a digital enterprise management data analysis system. The specific process is as follows: Step 110 to Step 140.

[0014] Step 110: Collect the supply chain dataset of the target enterprise.

[0015] In this example, a target enterprise with a complex business structure, whose supply chain encompasses multiple business scenarios and processes, is used as an example. Supply chain datasets are collected using an internal data collection architecture built within the enterprise. This architecture connects multiple data sources and collects order interaction records from the enterprise's order information management system. This data, presented in text format, contains various order-related information, such as order ID, order time, product category, and product quantity.

[0016] Logistics node monitoring data is collected through monitoring equipment distributed across various logistics nodes, recording the real-time status of logistics nodes in the form of images, including logistics node identification, images of cargo storage status, and images of transportation vehicle status. The collection process strictly adheres to data collection standards and specifications to ensure that the collected data is accurate, complete, and representative.

[0017] Step 120: Extract features from the supply chain data set to obtain a supply chain trend prediction feature set and a supply chain anomaly detection feature set.

[0018] In some examples, the supply chain dataset includes text-based order interaction record data and image-based logistics node monitoring data. Step 120 includes:

[0019] Step 121: Perform time-series image feature extraction processing on the logistics node monitoring data in the image form, extract the cargo stacking density change characteristics, transportation vehicle stay time characteristics and ambient light intensity fluctuation characteristics of each logistics node, and generate a dynamic feature set of the logistics node.

[0020] For logistics node monitoring data in image form, an image feature extraction algorithm is applied to each logistics node. First, to extract the characteristics of cargo stacking density variation, the cargo storage status images at different times are analyzed. Based on the distribution of cargo in the image and the image pixel information, the proportion of the image area occupied by cargo is calculated to represent the cargo stacking density. The density values ​​at different times are recorded to form a sequence of cargo stacking density variations over time, which serves as the cargo stacking density variation characteristic. For example, for logistics node M, the number of pixels distributed in the cargo storage status image from time t1 to time t2 is analyzed. Combined with the total number of pixels in the image, the cargo stacking density at time t1 is calculated as D1, the cargo stacking density at time t2 is calculated as D2, and so on. A series of density values ​​are obtained to form the cargo stacking density variation characteristic.

[0021] For the vehicle dwell time feature, we analyze the vehicle status images frame by frame to identify the time points when the vehicle enters and leaves the logistics node. We then calculate the time interval between the two points as the vehicle's dwell time at the node. For example, in a sequence of vehicle status images at logistics node N, image recognition technology determines that vehicle A enters at time t3 and leaves at time t4. Therefore, the dwell time of vehicle A at logistics node N is t4 - t3. The dwell time of multiple vehicles at different times is recorded to form a vehicle dwell time feature.

[0022] For the ambient light intensity fluctuation characteristics, we use the lighting information in the image and a common light intensity calculation method to calculate the light intensity value for each frame and analyze its changes at different times. For example, we use a common light intensity calculation model to process the ambient image of logistics node P, obtaining the light intensity value L1 at time t5 and the light intensity value L2 at time t6. We analyze the changes in these light intensity values ​​over time to obtain the ambient light intensity fluctuation characteristics. These extracted characteristics of cargo stacking density changes, vehicle dwell time characteristics, and ambient light intensity fluctuation characteristics are integrated to generate a dynamic feature set for the logistics node.

[0023] Step 122: Perform semantic association analysis on the text-based order interaction record data to extract demand keyword distribution features, supplier response timeliness features, and contract clause conflict semantic features from the order text to generate an order interaction semantic feature set.

[0024] Semantic analysis algorithms are used to process text-based order interaction record data. For demand keyword distribution features, the order text is first segmented, breaking the text into individual words. Then, a keyword extraction algorithm is used to identify demand-related keywords, such as "product name" and "quantity." The frequency and location of these keywords in the order text are then counted to construct demand keyword distribution features. For example, for order text T, segmentation yields a series of words. A keyword extraction algorithm is used to determine the demand keywords as "product X," "quantity Y," and so on. The number and location of "product X" and "quantity Y" occurrences in the text are then counted to form demand keyword distribution features.

[0025] The supplier response timeliness feature is measured by analyzing the time information about the supplier's response in the order text, such as the order issuance time and the supplier's response time. The time difference between the two is calculated to measure the supplier's response timeliness. For example, if order O is issued at time t7 and the supplier responds at time t8, the supplier response timeliness is t8 - t7. The supplier response timeliness information for multiple orders is recorded to form the supplier response timeliness feature.

[0026] For the semantic features of contract clause conflicts, we perform semantic analysis on the contract clause-related content in the order text and utilize a pre-set contract clause semantic model to identify potential conflicting semantics within the text. For example, we compare the descriptions of key contract clauses such as price and delivery date in different order texts to analyze whether there are any semantic contradictions or inconsistencies. If there are different wordings of price clauses in the order text, we use the semantic model to determine whether there is a conflict. This conflicting semantic information is then organized to form a semantic feature of contract clause conflicts. We then integrate the extracted demand keyword distribution features, supplier response timeliness features, and contract clause conflict semantic features to generate a set of order interaction semantic features.

[0027] Step 123: Call the pre-trained supply chain trend prediction model, perform cross-modal feature fusion processing on the logistics node dynamic feature set and the order interaction semantic feature set, and generate the supply chain trend prediction feature set.

[0028] In an optional embodiment, step 123 includes:

[0029] Step 1231: Time-align the cargo stacking density change characteristics and the transportation vehicle stay time characteristics in the logistics node dynamic feature set to generate a first time series feature vector; semantically encode the demand keyword distribution characteristics and the supplier response timeliness characteristics in the order interaction semantic feature set to generate a second time series feature vector.

[0030] Optionally, for the cargo stacking density change characteristics and transportation vehicle stay time characteristics in the logistics node dynamic feature set, since they come from different monitoring dimensions, but both have time series characteristics. First, the time series corresponding to the cargo stacking density change characteristics and transportation vehicle stay time characteristics are timestamp matched. For example, the time series of the cargo stacking density change characteristics is [D1, D2, D3, ...], and the corresponding timestamps are [t1, t2, t3, ...], and the time series of the transportation vehicle stay time characteristics is [L1, L2, L3, ...], and the corresponding timestamps are [t1', t2', t3', ...]. By comparing and adjusting the timestamps, the two time series are aligned so that they have corresponding feature values ​​at the same time point. Then, the aligned cargo stacking density change characteristics and transportation vehicle stay time characteristics are combined in a certain order to form a first time series feature vector. For example, the characteristic values ​​of the change in cargo stacking density and the characteristic values ​​of the length of stay of transport vehicles are arranged in sequence to obtain a first time series characteristic vector V1 = [D1, L1, D2, L2, D3, L3, ...].

[0031] A commonly used semantic encoding algorithm is used for the demand keyword distribution features and supplier response timeliness features in the order interaction semantic feature set. For the demand keyword distribution features, each demand keyword is mapped into a high-dimensional semantic space, and the semantic information of the keyword is converted into a vector representation through a commonly used mapping function. For example, the demand keyword "product X" is mapped to vector S1 through mapping function F1, and "quantity Y" is mapped to vector S2 through mapping function F1, and so on. For the supplier response timeliness feature, its numerical information is normalized so that it has the same dimension and scale range as the semantic vector. For example, the supplier response timeliness value t is normalized through the normalization function G to obtain the normalized value t'. Then, the semantically encoded demand keyword distribution feature vector and the normalized supplier response timeliness feature value are combined to form a second time series feature vector V2. For example, V2 = [S1, t', S2, t",...].

[0032] Step 1232: Use the attention weight allocation mechanism to dynamically adjust the weights of the first time series feature vector and the second time series feature vector to generate a fused cross-modal time series feature vector; wherein, the dynamic weight adjustment process includes allocating different attention weight coefficients according to the matching degree between the timestamp of the logistics node monitoring data and the timestamp of the order interaction record.

[0033] Optionally, an attention weight allocation mechanism is used to fuse the first and second time series feature vectors. First, the degree of match between the timestamps of the logistics node monitoring data and the timestamps of the order interaction records is calculated. For example, the time difference calculation function H is used to calculate the time difference between the feature value of the logistics node monitoring data at time t and the feature value of the order interaction record at a similar time t'. The degree of match is then determined based on a preset threshold. If the time difference is less than the threshold, the match is high; otherwise, the match is low.

[0034] Different attention weight coefficients are assigned to the first time series feature vector and the second time series feature vector according to the degree of matching. For example, when the degree of matching is high, a higher weight coefficient w1 is assigned to the first time series feature vector, and a lower weight coefficient w2 is assigned to the second time series feature vector; when the degree of matching is low, the weight coefficients are assigned in the opposite way. Then, the first time series feature vector and the second time series feature vector are fused by weighted summation. For example, the fused cross-modal time series feature vector V=w1*V1+w2*V2, where V1 is the first time series feature vector and V2 is the second time series feature vector. Through this dynamic weight adjustment process, the fused cross-modal time series feature vector can better integrate the information of the two modalities.

[0035] Step 1233: Input the cross-modal time series feature vector into the supply chain trend prediction model, extract trend change patterns of different time granularities through a multi-layer time convolutional network, and generate the supply chain trend prediction feature set.

[0036] Optionally, the fused cross-modal time series feature vector is input into a pre-trained supply chain trend prediction model, which contains multiple layers of temporal convolutional networks, each layer of which has a commonly used convolution kernel size and step size. First, the input cross-modal time series feature vector enters the first layer of temporal convolutional network, and the convolution kernel performs a sliding convolution operation on the feature vector in the time dimension. For example, the convolution kernel size is k, the step size is s, and the convolution kernel starts from the starting position of the feature vector and moves s positions each time, performing convolution calculations on feature vector segments of length k. Through convolution calculations, local features and trend change information at this time granularity are extracted.

[0037] An exemplary calculation process is as follows: for each position i in the feature vector V, the convolution calculation result C(i) = ∑(j = 0 to k - 1) w(j) * V(i + j), where w(j) is the weight coefficient of the convolution kernel. The output of the first layer of convolutional network serves as the input of the second layer of temporal convolutional network. The second layer of temporal convolutional network repeats the above convolution operation, but the convolution kernel size and step size may be different to further extract trend change patterns at more complex and macroscopic time granularities. Through layer-by-layer processing of multi-layer temporal convolutional networks, trend change patterns at different time granularities are ultimately extracted. These patterns constitute the supply chain trend prediction feature set.

[0038] Step 124: Use an anomaly detection convolutional network to perform local anomaly pattern recognition processing on the dynamic feature set of the logistics node to generate the supply chain anomaly detection feature set; the supply chain anomaly detection feature set includes cargo detention risk features, transportation route deviation features, and order matching anomaly features.

[0039] In an optional embodiment, step 124 includes:

[0040] Step 1241: Perform light anomaly detection processing on the ambient light intensity fluctuation characteristics in the dynamic feature set of the logistics node, extract the occurrence time and duration of the light intensity mutation event, and generate a light anomaly event feature; perform density gradient analysis processing on the cargo stacking density change characteristics, identify abnormal stacking areas where the density change rate exceeds a preset threshold, and generate a cargo detention risk feature; perform stay pattern clustering processing on the transport vehicle stay time characteristics, mark the transport vehicle that deviates from the preset stay time as a route deviation candidate, and generate a transport route deviation feature.

[0041] It can be understood that the ambient light intensity fluctuation characteristics are processed by the light anomaly detection algorithm. First, the rate of change of the light intensity is calculated by comparing the light intensity values ​​at adjacent moments, such as the light intensity at moment t is L(t), and the light intensity at moment t+1 is L(t+1), and the light intensity change rate R = |L(t+1)-L(t)| is calculated. When the light intensity change rate R exceeds the preset light intensity change threshold T1, it is determined that a light intensity mutation event has occurred. The occurrence time t and duration of the mutation event are recorded. The duration is obtained by calculating the time interval until the light intensity change rate returns to below the threshold, thereby generating the light anomaly event feature.

[0042] Density gradient analysis is performed to analyze the changing characteristics of cargo bulk density. This involves calculating the spatial and temporal gradient of cargo bulk density, such as the density variation at different locations and times within a logistics node. For the cargo bulk density D(x, t) at location x and time t, the density gradient is calculated by calculating the density difference between adjacent locations and times. When the density change rate exceeds a preset density change threshold T2, the area is identified as an abnormal accumulation area, marked, and its associated time information is used to generate a cargo detention risk profile.

[0043] To characterize vehicle dwell time, a dwell pattern clustering algorithm is used. First, a preset dwell time range is determined based on historical data or pre-set standards. Then, the current vehicle dwell time is analyzed and compared with the preset dwell time range. Vehicles whose dwell times deviate from the preset range are marked as route deviation candidates. Their identification, dwell time, and relevant logistics node information are recorded to generate a transport route deviation feature.

[0044] Step 1242: Correlate and match the lighting anomaly event feature, the cargo detention risk feature, and the transportation route deviation feature with the contract clause conflict semantic feature in the order interaction semantic feature set to generate the supply chain anomaly detection feature set.

[0045] Optionally, the lighting anomaly event features, cargo detention risk features, and transportation route deviation features are associated and matched with the contract terms conflict semantic features in the order interaction semantic feature set. By establishing association rules, for example, when an abnormal accumulation area in the cargo detention risk feature is detected to be associated with the contract terms related to the delivery of a certain order in the order interaction semantic feature set, they are matched. For the lighting anomaly event feature, if its occurrence time overlaps with the transportation time of a certain order, and the order in the order interaction semantic feature set has the contract terms conflict semantic feature, they are associated. For the transportation route deviation feature, if the goods transported by the deviated transportation vehicle are related to a certain order, and the order has the contract terms conflict semantic feature, it is also associated and matched. These associated and matched features are integrated to generate the supply chain anomaly detection feature set.

[0046] In an alternative embodiment, after generating the supply chain anomaly detection feature set, the method further includes:

[0047] Step 125: Input the cargo detention risk features and transportation route deviation features in the supply chain anomaly detection feature set into the reinforcement learning decision optimization model to trigger the anomaly feedback learning mechanism of the reinforcement learning decision optimization model; dynamically adjust the reward and punishment function weights in the anomaly feedback learning mechanism based on the historical processing records of the cargo detention risk features and the severity of the current transportation route deviation features; perform incremental training on the policy network of the reinforcement learning decision optimization model based on the adjusted reward and punishment function weights to generate an updated reinforcement learning decision optimization model; apply the updated reinforcement learning decision optimization model to subsequent supply chain anomaly detection feature set processing to achieve dynamic adaptive adjustment of the multi-stage decision optimization path.

[0048] In this embodiment of the present invention, cargo detention risk features and transportation route deviation features from the supply chain anomaly detection feature set are input into a reinforcement learning decision optimization model. Upon receiving these features, the model triggers an anomaly feedback learning mechanism. First, the model queries historical processing records for cargo detention risk features, which contain past decisions and outcomes for similar cargo detention risk situations. Simultaneously, the model assesses the severity of the current transportation route deviation feature, for example, by comprehensively evaluating factors such as the distance and duration of the deviation.

[0049] The reward and penalty function weights in the anomaly feedback learning mechanism are dynamically adjusted based on historical processing records and current severity. If historical processing records show that a certain decision-making method has achieved good results and the current transportation route deviation is severe, the corresponding reward function weight is increased. Conversely, if the historical decision-making results are poor and the current situation is relatively mild, the reward function weight is reduced or the penalty function weight is increased.

[0050] Based on the adjusted reward and penalty function weights, the policy network of the reinforcement learning decision optimization model undergoes incremental training. During training, the model adjusts the policy network parameters based on the current input features and the adjusted reward and penalty function weights. For example, using a gradient descent algorithm, the error between the policy network output and the expected output is calculated. The policy network weight parameters are then updated based on this error backpropagation, enabling the model to better respond to current anomalies. After multiple rounds of incremental training, an updated reinforcement learning decision optimization model is generated. This updated model is then applied to subsequent supply chain anomaly detection feature sets. The model then makes more informed decisions based on the new anomaly features, enabling dynamic adaptive adjustment of the multi-stage decision optimization path.

[0051] Step 130: Based on a preset reinforcement learning decision optimization model, dynamic decision path generation processing is performed on the supply chain trend prediction feature set and the supply chain anomaly detection feature set to generate a multi-stage decision optimization path.

[0052] In another example, the multi-stage decision optimization path is used to indicate the priority sequence of supply chain resource allocation and the execution order of exception handling operations. Step 130 includes:

[0053] Step 131: performing decision priority sorting on the target goods demand distribution features and logistics efficiency bottleneck prediction features in the supply chain trend prediction feature set to generate a resource allocation priority sequence.

[0054] In an embodiment of the present invention, the target goods demand distribution characteristics and logistics efficiency bottleneck prediction characteristics within the supply chain trend prediction feature set are prioritized for decision making. For the target goods demand distribution characteristics, factors such as the demand frequency and urgency of demand for different goods are analyzed. For example, for goods A and B, the urgency of demand is comprehensively assessed by statistically analyzing the demand frequency of goods A and B in historical order data, as well as the delivery deadlines for goods A and B in current orders. For the logistics efficiency bottleneck prediction characteristics, the location and impact of potential bottlenecks in the logistics process are analyzed, such as the potential for congestion on transportation routes and warehouse storage capacity limitations.

[0055] Based on these analysis results, a prioritization algorithm is employed, for example, based on demand urgency and bottleneck impact. If the demand for Good A is more urgent than Good B, and the impact of the bottleneck in the logistics link involved in Good A is relatively low, then Good A will be prioritized over Good B. The demand distribution characteristics of the sorted target goods and the predicted characteristics of logistics efficiency bottlenecks are integrated to generate a resource allocation priority sequence, clarifying the priority order of different goods and logistics links in supply chain resource allocation.

[0056] Step 132: Perform an abnormal impact assessment on the cargo detention risk features and transportation route deviation features in the supply chain abnormality detection feature set to generate an abnormality handling operation sequence.

[0057] Optionally, an abnormal impact assessment is performed on the cargo detention risk features and transportation route deviation features in the supply chain anomaly detection feature set. For the cargo detention risk feature, the impact of cargo detention on production progress, delivery deadlines, and other aspects is assessed. For example, the criticality of the detained cargo in the production process is analyzed. If the detained cargo is a key raw material for the production of a certain product and the delivery deadline of the product is approaching, the impact of the cargo detention is high. For the transportation route deviation feature, the impact of the deviation on transportation cost and transportation time is assessed, such as whether the deviation leads to a significant increase in transportation distance or extended transportation time.

[0058] Based on the assessment results, cargo detention risk characteristics and transport route deviation characteristics are ranked from high to low in terms of impact. The exception handling operations corresponding to the ranked characteristics are organized into an exception handling operation sequence, clarifying the order in which different operations should be executed when handling supply chain anomalies.

[0059] Step 133: Use a reinforcement learning decision optimization model to perform dynamic path combination optimization processing on the resource allocation priority sequence and the exception handling operation sequence to generate an initial multi-stage decision optimization path containing multiple decision sub-paths; wherein each decision sub-path corresponds to an optimized operation combination of a supply chain link.

[0060] The reinforcement learning decision optimization model performs dynamic path combination optimization based on the resource allocation priority sequence and the exception handling operation sequence. The model first comprehensively analyzes the resource requirements in the resource allocation priority sequence and the exception conditions in the exception handling operation sequence. For example, it considers the interrelationships and impacts between the demand for different goods in the resource allocation priority sequence, the potential problems identified by the logistics efficiency bottleneck prediction features, and the abnormal conditions in the exception handling operation sequence, such as the risk of cargo detention and transportation route deviation.

[0061] The model continuously tries different combinations, using its internal policy network to generate multiple decision subpaths. For example, in a supply chain involving cargo transportation and warehouse storage, the model comprehensively considers resource allocation priorities when generating decision subpaths. For example, if a high-priority shipment requires rapid transport to a specific warehouse, and the logistics node where the warehouse is located experiences an anomaly resulting in a transport route deviation, the model will use reinforcement learning strategies to try different combinations of transport solutions and exception handling measures.

[0062] For example, one combination might prioritize allocating a specific transportation vehicle to transport a particular shipment, while also temporarily creating an alternate route to ensure timely delivery to the warehouse in the event of a shipping route deviation. These combinations of operations, encompassing resource allocation and exception handling for different supply chain links, form decision subpaths. These subpaths are then integrated to generate an initial multi-stage decision optimization path encompassing multiple subpaths.

[0063] Step 134: According to the preset supply chain link weight distribution table, simulate and predict the execution effect of each decision sub-path, and select the decision sub-path combination with the best prediction effect as the multi-stage decision optimization path.

[0064] Optionally, a preset supply chain link weighting table specifies the importance of different supply chain links in the overall supply chain operation. For example, the production link has a weight of w1, the transportation link has a weight of w2, and the warehousing link has a weight of w3. Based on this weighting table, a simulation is performed to predict the execution effect of each decision subpath.

[0065] For each decision subpath, the simulation and forecasting process considers the supply chain links involved in that subpath and their impact on the overall supply chain objectives. For example, a decision subpath involves the transportation link (transporting goods from the production site to the warehouse) and the warehousing link (arranging the storage of the goods). The simulation and forecasting process calculates a comprehensive score based on the weight w2 of the transportation link and the weight w3 of the warehousing link, combined with factors such as the transportation cost and time of the transportation solution in that decision subpath, and the storage cost and cargo loss of the warehousing solution.

[0066] For example, through a commonly used evaluation function F, various indicators of the transportation link such as transportation cost C1, transportation time T1, and various indicators of the warehousing link such as storage cost C2, cargo loss rate L, etc. are calculated with the weights of the corresponding links: score S = w2*F1(C1, T1)+w3*F2(C2, L), where F1 and F2 are specific evaluation functions for different indicators of the transportation link and warehousing link.

[0067] After performing this simulation and prediction calculation on all decision subpaths, the scores are compared and the highest-scoring combination of subpaths (i.e., the one with the best prediction results) is selected as the multi-stage decision optimization path. This multi-stage decision optimization path maximizes overall supply chain operations, taking into account the importance of different supply chain links, ensuring rational resource allocation and effective handling of abnormal situations.

[0068] Step 140: Generate a visual graphic analysis report based on the multi-stage decision optimization path and the supply chain data set, and feed the visual graphic analysis report back to the enterprise management system; wherein the visual graphic analysis report is used to instruct the enterprise management system to perform supply chain dynamic adjustment operations.

[0069] In a preferred embodiment, step 140 includes:

[0070] Step 141: Convert the resource configuration priority sequence in the multi-stage decision optimization path into a first visual decision tree structure, and mark the demand satisfaction rate and resource consumption rate of each resource configuration node in the first visual decision tree structure.

[0071] First, the resource allocation priority sequence in the multi-stage decision optimization path is structured and converted into a first visual decision tree structure. Starting with the highest-priority resource in the resource allocation priority sequence, this resource is used as the root node of the decision tree. For example, if resource A ranks first in the resource allocation priority sequence, then resource A is the root node.

[0072] For each resource node, child nodes are generated based on its subsequent resource allocation decisions and dependencies. For example, if the allocation of resource A may depend on transportation resource B and storage resource C, then transportation resource B and storage resource C can be child nodes of resource A. In this way, a complete decision tree structure is gradually constructed.

[0073] Next, calculate and annotate the demand fulfillment rate and resource consumption rate for each resource configuration node. The demand fulfillment rate is calculated by taking the ratio of the actual quantity allocated to the resource to the expected quantity demanded. For example, for resource D, the expected quantity demanded is N1, and the actual quantity allocated is N2. The demand fulfillment rate = N2 / N1. The resource consumption rate is calculated by taking the ratio of the actual quantity consumed during the allocation and use of the resource to the total available quantity. For example, if the total available quantity of resource E is M1, and the actual quantity consumed is M2, the resource consumption rate = M2 / M1. These calculated demand fulfillment rates and resource consumption rates are annotated on the corresponding resource configuration nodes to form the first visual decision tree structure, which intuitively displays the priority of resource configuration and related indicators.

[0074] Step 142: Convert the exception handling operation sequence in the multi-stage decision optimization path into a second visual decision tree structure, and annotate the risk reduction rate and operation execution cost of each exception handling node in the second visual decision tree structure.

[0075] Optionally, the sequence of exception handling operations in the multi-stage decision optimization path is processed to transform it into a second visual decision tree structure. The starting point of the exception handling process, for example, the operation addressing the cargo detention risk, serves as the root node. Child nodes are generated sequentially based on the exception handling process and subsequent operations. If addressing the cargo detention risk requires an inventory count followed by a transport reallocation operation, the inventory count and transport reallocation operations can serve as child nodes of the cargo detention risk handling root node.

[0076] For each exception handling node, calculate and mark the risk reduction rate and operation execution cost. The risk reduction rate is calculated by the change in risk assessment value before and after processing. For example, before processing the transportation route deviation exception, the risk value of the exception to the supply chain is assessed to be R1, and the risk value is reduced to R2 after processing, then the risk reduction rate = (R1-R2) / R1. The operation execution cost includes the sum of all costs related to the execution of the operation, such as labor costs and material costs. For example, the labor cost of executing a certain exception handling operation is C3, the material procurement cost is C4, etc., then the operation execution cost = C3+C4. The risk reduction rate and operation execution cost are marked on the corresponding exception handling node to form a second visual decision tree structure, which clearly presents information related to the process and effect of exception handling.

[0077] Step 143: Perform key frame extraction processing on the image-shaped logistics node monitoring data in the supply chain data set to generate a logistics status change animation, and perform spatiotemporal alignment processing on the logistics status change animation with the first visual decision tree structure and the second visual decision tree structure.

[0078] For example, a keyframe extraction algorithm can be used to determine keyframes based on the degree of change in image content. This can be done by comparing factors such as pixel differences between adjacent image frames and the motion of objects. If the stacking of cargo changes significantly between adjacent image frames, or if the position of a transport vehicle shifts significantly, that frame may be selected as a keyframe.

[0079] After extracting key frames from the image sequence of logistics node monitoring data, these key frames are arranged in chronological order and appropriate transition effects are added to generate a logistics status change animation. This animation can intuitively show the status changes of logistics nodes over a period of time, such as the entry and exit of goods, the stop and departure of transportation vehicles, etc.

[0080] Next, the logistics status change animation is spatially and temporally aligned with the first and second visual decision tree structures. Based on temporal information, the time points in the animation are matched with the time-related information in the decision tree structures. For example, the execution time of a resource allocation operation in the first visual decision tree structure corresponds to the time when the goods are transported to the corresponding warehouse in the logistics status change animation; and the time of an exception handling operation in the second visual decision tree structure is associated with the time when the exception occurs in the logistics status change animation. This spatial and temporal alignment establishes a close connection between the logistics status change animation and the two visual decision tree structures, facilitating subsequent comprehensive display and analysis.

[0081] Step 144: Interactively layout matching processing is performed on the aligned visual decision tree structure and the logistics status change animation to generate the visual graphic analysis report containing the dynamic data association relationship.

[0082] Optionally, interactive layout matching is performed on the aligned first and second visual decision tree structures and the logistics status change animation. First, an overall layout framework is determined, and the two visual decision tree structures and the logistics status change animation are rationally arranged within a single display interface. For example, the first visual decision tree structure can be placed on the left side of the interface, the second visual decision tree structure on the right side, and the logistics status change animation in the center.

[0083] Then, an interactive association is established. When the user hovers over a resource configuration node in the first visual decision tree structure, the corresponding logistics operation is highlighted in the logistics status change animation. Simultaneously, the exception handling node related to that resource configuration in the second visual decision tree structure is also highlighted. For example, when the user hovers over the configuration node for resource A, the logistics status change animation displays the transportation process of resource A, and the second visual decision tree structure displays the processing node if an exception occurs in the transportation of resource A.

[0084] Conversely, when a user operates an exception handling node in the second visual decision tree structure, the resource allocation nodes affected by the exception in the first visual decision tree structure, as well as the corresponding exception occurrence time and handling process in the logistics status change animation, are also displayed. This interactive layout matching process generates a visual graphic analysis report containing dynamic data associations, providing business managers with an intuitive and comprehensive display of supply chain information, facilitating decision-making and dynamic supply chain adjustment operations.

[0085] In another alternative embodiment, after generating the visual graphic and text analysis report, the process further includes:

[0086] Step 145: Monitor the response operation data of the enterprise management system to the visual graphic analysis report, extract the resource configuration adjustment data and exception handling execution data in the response operation data; perform difference comparison processing on the resource configuration adjustment data and the resource configuration priority sequence in the multi-stage decision optimization path to generate a resource configuration policy offset feature; perform effect evaluation processing on the exception handling execution data and the exception handling operation sequence in the multi-stage decision optimization path to generate an exception handling effect lag feature; feed back the resource configuration policy offset feature and the exception handling effect lag feature to the reinforcement learning decision optimization model to trigger the online strategy adjustment mechanism of the reinforcement learning decision optimization model; update the multi-stage decision optimization path according to the output result of the online strategy adjustment mechanism, and regenerate the visual graphic optimization report containing the updated decision path.

[0087] Optionally, use common monitoring tools to monitor the enterprise management system's response data to the visual graphical analysis reports in real time. From this response data, accurately extract resource configuration adjustment data and exception handling execution data. For example, resource configuration adjustment data may include new resource allocation plans, changes in resource priorities, etc.; exception handling execution data may include the actual exception handling operations performed and the execution time of the operations.

[0088] Compare the resource allocation adjustment data with the resource allocation priority sequence in the multi-stage decision optimization path. Calculate the difference between the actual resource allocation and the originally planned resource allocation priority. For example, if the original plan prioritized resource A over resource B, but the actual resource allocation allocated more resource B than resource A, calculate the degree and direction of this difference to generate a resource allocation strategy shift feature.

[0089] Exception handling execution data is evaluated against the exception handling operation sequence in the multi-stage decision optimization path. The risk reduction effect achieved by the actual exception handling operations is compared with the expected effect. If the actual risk reduction rate is lower than the expected risk reduction rate, it indicates that there is a lag in the exception handling effect. By quantifying this lag, a lag feature is generated.

[0090] The resource allocation policy deviation characteristics and the lag characteristics of exception handling effectiveness are fed back into the reinforcement learning decision optimization model. After receiving this feedback, the model triggers an online policy adjustment mechanism, which adjusts the model's policy network based on the feedback characteristics. For example, if the resource allocation policy deviation is large, the model increases the reward weight associated with correct resource allocation; if the exception handling effect is lagging, the model adjusts the exception handling policy parameters to improve the effectiveness of future exception handling.

[0091] Based on the output of the online policy adjustment mechanism, the multi-stage decision optimization path is updated. For example, resource allocation priorities can be adjusted, and the execution order of exception handling operations can be modified. Then, based on the updated multi-stage decision optimization path, a visual optimization report containing the updated decision path is regenerated, providing more accurate and effective decision support information for the enterprise management system.

[0092] In one implementation, the training process of the preset reinforcement learning decision optimization model includes:

[0093] Step 210: Obtain a historical supply chain dataset and corresponding historical decision path execution effect data.

[0094] For example, a historical supply chain dataset is obtained from the company's historical data storage system. This dataset contains various supply chain-related information, such as order interaction records and logistics node monitoring data from the past period. Simultaneously, historical decision path execution performance data corresponding to this historical data is collected. This data records the decision paths made in the past based on supply chain conditions and the evaluation of the actual performance of these decision paths after execution. For example, information such as whether resource allocation under a specific decision path is reasonable and whether abnormal situations are effectively handled is recorded. This historical data and execution performance data provide a rich sample and reference basis for subsequent model training.

[0095] Step 220: Perform feature enhancement processing on the historical supply chain data set to generate a simulated supply chain trend prediction feature set and a simulated supply chain anomaly detection feature set.

[0096] Optionally, feature enhancement processing can be performed on historical supply chain datasets to enrich the data's feature information. For order interaction records, in addition to extracting conventional demand keyword distribution features and supplier response timeliness features, text mining techniques can also be used to extract potential market trends and customer preference features from order text. For example, by analyzing customer descriptions of product features in order text, potential customer preferences can be discovered.

[0097] For logistics node monitoring data, in addition to the original characteristics of changes in cargo stacking density and transportation vehicle stay time, image processing technology is used to enhance the detailed information of the image and extract more detailed logistics operation characteristics, such as cargo handling method characteristics and transportation vehicle loading and unloading efficiency characteristics.

[0098] The enhanced order interaction record data and logistics node monitoring data are integrated and classified to generate a simulated supply chain trend prediction feature set and a simulated supply chain anomaly detection feature set. The simulated supply chain trend prediction feature set includes various features that can be used to predict future supply chain trends, while the simulated supply chain anomaly detection feature set includes various features for detecting supply chain anomalies. These feature sets provide more comprehensive and valuable data support for model training.

[0099] Step 230: Construct an initial reinforcement learning decision optimization model and set a compound reward function that includes resource allocation efficiency rewards and exception handling effect rewards.

[0100] The model uses a network structure suitable for reinforcement learning, such as the Deep Q-Network (DQN) or its variants. The input of the model is a simulated supply chain trend prediction feature set and a simulated supply chain anomaly detection feature set, and the output is the action selection for supply chain decision-making.

[0101] Set up a composite reward function that includes rewards for resource allocation efficiency and exception handling effectiveness. Resource allocation efficiency rewards are measured based on the rationality and effectiveness of resource allocation. For example, if the model's resource allocation decisions increase resource demand satisfaction while reducing resource consumption, a higher reward will be given. Conversely, if resource allocation is unreasonable, resulting in resource waste or unmet demand, a lower reward or penalty will be given.

[0102] Exception handling effectiveness rewards are evaluated based on the actual results of exception handling. If the exception handling operation successfully reduces risk and the execution cost is within a reasonable range, a corresponding reward is given; if the exception handling operation fails to effectively reduce risk or the cost is too high, negative feedback is given. The composite reward function combines these two aspects of rewards, allowing the model to simultaneously consider resource allocation and exception handling effectiveness during training, thereby learning more optimized decision-making strategies.

[0103] Step 240: Input the simulated supply chain trend prediction feature set and the simulated supply chain anomaly detection feature set into the initial reinforcement learning decision optimization model, and generate the reinforcement learning decision optimization model that meets the preset decision effect through multiple rounds of strategy exploration and strategy utilization alternating training; wherein, the strategy exploration stage includes randomly attempting decision sub-paths that have not appeared in the historical decision path execution effect data, and the strategy utilization stage includes selecting the historical optimal decision sub-path for reinforcement learning based on the feedback value of the compound reward function.

[0104] For example, the simulated supply chain trend prediction feature set and the simulated supply chain anomaly detection feature set are input into the initial reinforcement learning decision optimization model for training. During the training process, multiple rounds of strategy exploration and strategy utilization are performed alternately.

[0105] During the strategy exploration phase, the model randomly attempts decision subpaths not previously seen in historical decision path execution data. For example, for resource allocation decisions, it might try a new resource allocation combination; for exception handling decisions, it might explore a different exception handling process. Through this random experimentation, the model can discover new and potentially effective decision-making approaches, expanding the decision space.

[0106] During the strategy implementation phase, the model selects the historically optimal decision subpath for reinforcement learning based on the feedback from the compound reward function. If a decision subpath has received a high reward in the past, the model increases the probability of selecting that decision subpath, further reinforcing this successful decision strategy.

[0107] Through multiple rounds of this alternating training, the model continuously adjusts its policy network parameters to adapt to different supply chain situations, gradually learning more optimized decision-making strategies. When the model's decision-making results meet preset criteria, such as a certain level of resource allocation efficiency or a certain percentage of exception handling success rate, a reinforcement learning decision optimization model that meets the preset decision-making results is generated.

[0108] In an independent embodiment, after the visual graphic analysis report is fed back to the enterprise management system, it also includes: performing key area enhancement processing on the logistics status change animation in the visual graphic analysis report, extracting the frame image in which the cargo stacking density exceeds the threshold in the animation, and generating a highlighted area; performing spatial coordinate mapping processing on the highlighted area and the resource allocation node in the first visual decision tree structure to generate an enhanced decision tree layer with a heat map superimposed; dynamically adjusting the transparency of the enhanced decision tree layer, and dynamically adjusting the transparency parameter of the corresponding heat map according to the priority of the resource allocation node; performing frame synchronization fusion processing on the adjusted enhanced decision tree layer and the logistics status change animation to generate a dynamic interactive enhanced report and update it to the enterprise management system.

[0109] In this embodiment, key areas of the logistics status change animation in the visual graphic analysis report are enhanced. First, a threshold for cargo density is set. The cargo density is calculated by analyzing each frame of the animation. For example, using image pixel analysis technology, the cargo density is expressed by calculating the ratio of the number of pixels occupied by cargo to the total number of pixels in the image. When the cargo density of a frame exceeds the set threshold, that frame is extracted and the images are combined to generate a highlighted area.

[0110] Perform spatial coordinate mapping between the highlighted area and the resource allocation node in the first visual decision tree structure. First, determine the spatial correspondence between the logistics scene in the animation and the resource allocation scene represented by the first visual decision tree structure. For example, the logistics warehouse in the animation has a common location marker in real space, and the resource allocation node in the first visual decision tree structure also has a logical association with the warehouse.

[0111] By establishing this spatial correspondence, each pixel position in the highlighted area is mapped to the spatial position of a resource allocation node. For example, if a location in the highlighted area represents a specific area of ​​accumulated goods in a warehouse, this location is mapped to the resource allocation node associated with that warehouse in the first visual decision tree structure. The coordinate correspondence between them is determined, thereby generating an enhanced decision tree layer with an overlay of the heat map.

[0112] Next, the transparency of the enhanced decision tree layer is dynamically adjusted. The transparency parameters of the corresponding heat map are dynamically adjusted according to the priority of the resource configuration node. For example, for resource configuration nodes with higher priorities, the transparency of their corresponding heat maps is set higher to make them more prominent in the layer; for resource configuration nodes with lower priorities, the transparency of their heat maps is reduced accordingly. The specific implementation method can be through a transparency adjustment function, which takes the priority of the resource configuration node as an input parameter and outputs the corresponding transparency value. For example, the priority range is set to [1, 10]. The higher the priority, the larger the value. The transparency adjustment function is: transparency = priority / 10 (the value range is between 0 and 1). In this way, the transparency of the heat map can be dynamically adjusted according to different priorities.

[0113] Finally, the adjusted enhanced decision tree layer is frame-synchronized and fused with the logistics status change animation. During the fusion process, the time information of the enhanced decision tree layer is ensured to accurately match the time information of the logistics status change animation. For example, when the logistics status change animation plays to a certain moment, the corresponding enhanced decision tree layer also displays the relevant resource allocation and heat map information at that moment. Through this frame-synchronized fusion process, a dynamic interactive enhanced report is generated and updated to the enterprise management system. When enterprise managers view the visual graphic analysis report, they can more intuitively understand the relationship between areas of abnormal cargo stacking density and resource allocation, as well as the dynamic information that changes over time, allowing them to make more accurate decisions.

[0114] In an independent embodiment, after the visual graphic analysis report is fed back to the enterprise management system, it also includes: identifying the supply chain anomaly detection feature keywords in the text description segment of the visual graphic analysis report, and generating an abnormal position annotation box associated with each supply chain anomaly detection feature keyword; performing semantic matching processing on the abnormal position annotation box and the abnormal processing node of the second visual decision tree structure to generate an abnormal processing topology map with multi-layer annotation boxes; performing layout conflict detection processing on the multi-layer annotation boxes, automatically adjusting the annotation level distribution and folding and expanding trigger conditions according to the overlapping area of ​​the annotation boxes, and obtaining an adjusted abnormal processing topology map; performing cross-layer association processing on the adjusted abnormal processing topology map and the order interaction semantic feature set to generate an interactive report that supports click-triggered display of semantic association links and returns updates.

[0115] First, keyword recognition is performed on the text descriptions in the visualization graphic analysis report. Natural language processing techniques, including lexical and syntactic analysis, are used to identify keywords related to supply chain anomaly detection features. For example, keywords such as "cargo delay," "transportation route deviation," and "contract term conflict" are extracted from the text. For each keyword, an annotation box is generated that identifies the associated anomaly location. The annotation box can include a brief description of the anomaly detection feature represented by the keyword and information about the location where the anomaly may have occurred.

[0116] These anomaly location annotation boxes are then semantically matched with the anomaly handling nodes in the second visual decision tree structure. A semantic analysis algorithm compares the content of the annotation boxes with the semantic information of the anomaly handling nodes to determine their relevance. For example, if an anomaly location annotation box mentions "Cargo detention risk at Warehouse A," and the second visual decision tree structure contains an anomaly handling node specifically addressing cargo detention at Warehouse A, the two are matched. This matching generates an anomaly handling topology diagram with multiple layers of annotation boxes, making the anomaly handling process and related anomaly location information more clear and intuitive.

[0117] Next, layout conflict detection is performed on multi-layer annotation boxes. By calculating the positional relationship and overlapping areas between annotation boxes, a determination is made as to whether a layout conflict exists. For example, if two annotation boxes overlap, resulting in unclear information display, a layout conflict is considered to exist. Based on the detected conflict, the annotation hierarchy distribution and folding and expanding trigger conditions are automatically adjusted. For example, if an annotation box is covered by other annotation boxes, it is adjusted to a higher level for display. At the same time, a folding and expanding trigger condition is set so that when the user hovers over an annotation box, the related detailed information is expanded. Through these adjustments, an adjusted exception handling topology diagram is obtained, which improves the readability and display effect of the annotation information.

[0118] Finally, the adjusted exception handling topology map is cross-layer associated with the order interaction semantic feature set. By establishing semantic association rules, the exception information in the exception handling topology map is associated with the relevant semantic information in the order interaction semantic feature set. For example, if an exception handling node in the exception handling topology map involves a "contract terms conflict", the order interaction semantic feature set is searched for the relevant contract terms conflict semantic feature information and a link is established. An interactive report is generated that supports click-triggered display of semantic association links. When the user clicks on a node in the exception handling topology map, the user can view the related order interaction semantic information and further understand the root cause and impact of the exception. This interactive report is fed back and updated to the enterprise management system to provide enterprise managers with more comprehensive and in-depth supply chain exception analysis information, which helps them better formulate response strategies.

[0119] It should be noted that when implementing the above technical solution, technical personnel in the relevant field can dynamically align the heterogeneous timestamps of logistics node monitoring data and order data based on the time series interpolation algorithm in the existing technology (such as cubic spline interpolation), thereby solving the time synchronization problem of data with different sampling frequencies.

[0120] Regarding the attention weight allocation mechanism, multi-head attention technology can be used to jointly model time matching and feature importance, and the feature correlation matrix can be constructed by calculating the mutual information between logistics dynamic features and order semantic features, thereby optimizing the dynamic weight allocation strategy of cross-modal features.

[0121] For the setting of anomaly detection thresholds, an adaptive threshold calculation method based on historical data (such as the 3σ principle) can be introduced, combined with the sliding window statistics of the mean and variance of the cargo stacking density of each logistics node, and differentiated anomaly judgments of nodes of different sizes can be achieved by dynamically adjusting the threshold boundaries.

[0122] In terms of multi-model interface dimension matching, feature embedding technology (such as the BERT pre-training model) can be used to vectorize text features, while convolutional neural networks (CNNs) can be used to perform spatial dimensionality reduction on image features, and the feature vector dimensions can be unified through the fully connected layer to achieve compatible input of cross-modal data.

[0123] To address the dimensional conversion problem of cargo stacking density, a mapping relationship table between the pixel distribution of logistics monitoring images and the actual volume of the storage space can be established based on the FastDTW algorithm. Through calibration experiments, the pixel ratio-cubic volume conversion coefficient of typical cargo stacking scenarios can be measured to achieve quantitative conversion of pixel features to physical density. For time unit standardization, the UNIX timestamp mechanism can be used to synchronize the clock sources of the order system and logistics equipment at the millisecond level, and Z-score standardization can be used to eliminate the dimensional differences in the time recording accuracy of different business systems. In terms of unifying cost calculation units, the human cost accounting module of the enterprise ERP system can be integrated to convert the labor hours involved in exception handling operations into monetary units in real time according to job salary standards, and normalization of multi-dimensional cost indicators can be achieved through linear weighting.

[0124] To improve the quantification of light intensity, the brightness value of the V channel of the HSV color space of the monitoring image can be extracted based on the OpenCV image processing library. The brightness-illuminance regression model can be constructed in combination with the calibration data of the photometric sensor to achieve accurate mapping between pixel brightness values ​​and physical illuminance units (Lux).

[0125] In the dynamic decision path generation stage, the curriculum reinforcement learning (Curriculum Reinforcement Learning) technology can be introduced to improve the model's exploration efficiency of multi-stage decision combinations through a training strategy that gradually increases the complexity of the environment. At the same time, the Monte Carlo Tree Search (MCTS) algorithm is combined to conduct game deduction on the decision sub-paths, thereby optimizing the collaborative strategy of resource allocation and exception handling.

[0126] Through the combined application of the above-mentioned existing technologies, the logic and dimensions of the embodiments of the present invention can be systematically optimized to ensure a full-process technical closed loop of supply chain data from collection, feature extraction to decision generation.

[0127] In summary, the embodiments of the present invention are based on comprehensive and diverse supply chain data and integrate dynamic decision-making logic to generate visual graphic analysis reports, which can improve the comprehensiveness, dynamism and decision-making guidance of graphic generation, thereby effectively guiding the enterprise management system to dynamically adjust the supply chain.

[0128] In detail, the embodiments of the present invention provide a comprehensive and diverse data foundation for the generation of graphics and text by collecting supply chain data sets that include text and image formats; the supply chain trend prediction feature set and supply chain anomaly detection feature set obtained through feature extraction inject depth and pertinence into the graphics and text content; the multi-stage decision optimization path generated based on the reinforcement learning decision optimization model further clarifies the logical framework for graphics and text generation, ensuring that the generated visual graphics and text analysis report is not a simple list of data, but an in-depth analysis result with logical coherence and decision-making orientation. The visual graphics and text analysis report not only intuitively presents the supply chain status, but also instructs the enterprise management system to perform dynamic supply chain adjustment operations in a clear and understandable graphic form based on the multi-stage decision path, thereby improving the visualization and intelligence level of enterprise decision-making.

[0129] Based on the same inventive concept, the embodiment of the present invention also provides a digital enterprise management data analysis system. Figure 2 As shown, it is a structural diagram of a possible digital enterprise management data analysis system provided in an embodiment of the present invention. Figure 2 In the embodiment, the digital enterprise management data analysis system 200 includes a processor 210 and a memory 220. The memory 220 stores a computer program executable by the processor 210. The processor 210 can perform the steps of the above-mentioned digital enterprise management data analysis method based on artificial intelligence by executing the instructions stored in the memory 220.

[0130] Based on the same inventive concept, an embodiment of the present invention provides a computer-readable storage medium, which includes a computer program. When the computer program is run on a digital enterprise management data analysis system, the computer program is used to enable the digital enterprise management data analysis system to perform the steps of the digital enterprise management data analysis method based on artificial intelligence. In some possible implementations, the various aspects of the digital enterprise management data analysis method based on artificial intelligence provided by the present invention can also be implemented in the form of a program product, which includes a computer program. When the program product is run on a digital enterprise management data analysis system, the computer program is used to enable the digital enterprise management data analysis system to perform the steps of the digital enterprise management data analysis method based on artificial intelligence. For example, the digital enterprise management data analysis system can perform the following steps: Figure 1 Follow the steps shown in .

[0131] In the technical solutions involved in the above-mentioned embodiments of the present invention, whether it is performing comparison calculations of multi-dimensional features or constructing composite parameters, if there are problems caused by significant differences in the number of dimensions, dimensional units and semantic meanings of different features, technical personnel in this field, based on their professional knowledge and past practical experience, are fully able to understand that these differences need to be properly handled so that the calculation results are accurate and comparable, and avoid situations such as logical confusion and unclear mathematical meaning.

[0132] Specifically, when faced with features having different numbers of dimensions, those skilled in the art may employ various strategies to accurately calculate the similarity, matching degree, or feature distance between different features.

[0133] Feature selection is a common method. For a high-dimensional feature set, we can select the most representative subset of features from the high-dimensional features, matching the number of low-dimensional features based on metrics such as feature importance and relevance. Feature selection is performed using methods such as the chi-square test and information gain to identify the most valuable features for the technical solution. This reduces the high-dimensional features to a dimension comparable to the low-dimensional features, allowing for similarity or distance calculations.

[0134] Feature extraction is also an effective method. By constructing a suitable feature extraction model, features of different dimensions can be mapped into a common low-dimensional feature space. Principal component analysis (PCA) can not only be used to handle dimensional differences, but also project high-dimensional features into a low-dimensional space composed of principal components, making features of different dimensions comparable in this low-dimensional space. In addition, deep learning models such as autoencoders can also be used for feature extraction. They can automatically learn the latent representation of input features and convert features of different dimensions into feature vectors of the same dimension for subsequent similarity, matching, or feature distance calculations.

[0135] Alternatively, kernel methods can be used. Kernel functions can calculate the similarity between features in a high-dimensional space without explicitly mapping the features to that space. For features with varying numbers of dimensions, appropriate kernel functions, such as Gaussian or polynomial kernel functions, can be selected to directly calculate the similarity between them. This approach avoids the direct computational difficulties associated with varying feature dimensions and effectively measures the relationships between features in either the original feature space or an implicit high-dimensional space.

[0136] When processing the comparison of multi-dimensional features, in order to achieve comparable alignment of feature spaces, those skilled in the art may adopt a variety of existing common technical means.

[0137] Standardization preprocessing is a widely used and effective method. It transforms raw feature data into a standard normal distribution with a mean of 0 and a standard deviation of 1 by performing a specific linear transformation on the data. This process essentially eliminates the influence of different dimensions between features, allowing all features to be compared at the same scale. For example, in a dataset containing features of different dimensions, after standardization preprocessing, similarity or distance calculations can be performed on these features at the same scale, avoiding calculation bias caused by different dimensions.

[0138] Mapping transformation is also an effective way to solve the problem of dimensional differences. It can map the original features into a new space based on the specific properties of the features and actual business needs. In this new space, features of different dimensions can be better comparable. For those features with nonlinear relationships, those skilled in the art can use logarithmic transformation, power transformation, etc. to convert them into linear relationships, which makes it easier to calculate similarity or distance. For example, when processing some features with exponential growth trends, they can be converted into linear relationships through logarithmic transformation, making subsequent calculations more accurate and convenient.

[0139] Spatial projection is also an important technical means. It projects high-dimensional feature space into low-dimensional space while retaining important information between features as much as possible. By carefully selecting the appropriate projection direction and projection dimension, technicians in this field can effectively reduce the impact of dimensional differences on calculation results while reducing the data dimension. Common spatial projection methods include principal component analysis (PCA) and linear discriminant analysis (LDA). Taking principal component analysis as an example, it projects high-dimensional data into a low-dimensional space composed of principal components by finding the principal component direction of the data. While simplifying the data structure, it also reduces the interference of dimensional differences on feature comparison.

[0140] In the process of constructing composite parameters (such as loss function values), different parameter items often have different dimensions. Those skilled in the art can adopt normalization processing or an adaptive weight allocation mechanism based on distribution characteristics.

[0141] Normalization involves unifying the value ranges of different parameter items into a fixed interval, such as [0, 1]. This approach eliminates the impact of dimensional differences and ensures that all parameter items have equal importance during weighted fusion. Common normalization methods include min-max normalization and Z-score normalization. For example, min-max normalization performs a linear transformation on parameter items, scaling their value range to the interval [0, 1]. This allows weighted fusion of parameters of different dimensions to be performed under the same standard.

[0142] The adaptive weight allocation mechanism based on distribution characteristics dynamically adjusts the weights of different parameter items according to their distribution characteristics. For parameter items with large variance, those skilled in the art can appropriately reduce their weights; for parameter items with small variance, those skilled in the art can appropriately increase their weights. Doing so can make the composite loss function pay more attention to those parameter items with smaller variances, thereby improving the stability and generalization ability of the model. For example, in a composite loss function containing multiple parameter items, if the variance of a parameter item is large, it means that its fluctuation is more drastic, which may have an adverse effect on the stability of the model. In this case, reducing its weight can reduce this adverse effect; and for parameter items with smaller variance, increasing their weight can make the model pay more attention to the information reflected by the parameter item, thereby improving the overall performance of the model.

[0143] The general technical approaches described above for solving the feature matching and loss balancing problems are common knowledge in the field. These techniques have been fully validated and widely used in numerous practical applications, and those skilled in the art can skillfully and flexibly apply these methods to address similar dimensional discrepancies.

[0144] The formulas and calculation processes involved in the embodiments of the present invention, whether used for multi-dimensional feature comparison or composite loss function construction, strictly follow the principle of dimensional correspondence. The variables in each formula have clear and definite physical meanings, and their operation logic is also fully consistent with basic mathematical and physical logic. The operation results must be the reasonable results expected by the present invention. Those skilled in the art have the ability to comprehensively apply the above-mentioned general technical means according to specific data conditions and business needs, and effectively solve the various problems caused by the number of dimensions, dimensional differences, etc. in the multi-dimensional feature comparison calculation and composite loss function construction in the embodiments, and ensure the accuracy, reliability and feasibility of the technical solution of the present invention.

Claims

1. A digital enterprise management data analysis method based on artificial intelligence, characterized in that: include: Collect supply chain data sets of target enterprises; The supply chain data set includes order interaction record data in text form and logistics node monitoring data in image form; Extracting features from the supply chain data set to obtain a supply chain trend prediction feature set and a supply chain anomaly detection feature set; Based on a preset reinforcement learning decision optimization model, dynamic decision path generation processing is performed on the supply chain trend prediction feature set and the supply chain anomaly detection feature set to generate a multi-stage decision optimization path; A visual graphic analysis report is generated based on the multi-stage decision optimization path and the supply chain data set, and the visual graphic analysis report is fed back to the enterprise management system; wherein the visual graphic analysis report is used to instruct the enterprise management system to perform supply chain dynamic adjustment operations.

2. The method according to claim 1, wherein The feature extraction of the supply chain data set to obtain a supply chain trend prediction feature set and a supply chain anomaly detection feature set includes: Performing time-series image feature extraction processing on the logistics node monitoring data in the form of images, extracting the cargo stacking density variation characteristics, transportation vehicle dwelling time characteristics, and ambient light intensity fluctuation characteristics of each logistics node, and generating a dynamic feature set of the logistics node; Performing semantic association analysis on the text-based order interaction record data to extract demand keyword distribution features, supplier response timeliness features, and contract clause conflict semantic features from the order text to generate an order interaction semantic feature set; Calling a pre-trained supply chain trend prediction model to perform cross-modal feature fusion processing on the logistics node dynamic feature set and the order interaction semantic feature set to generate the supply chain trend prediction feature set; An anomaly detection convolutional network is used to perform local anomaly pattern recognition processing on the dynamic feature set of the logistics node to generate the supply chain anomaly detection feature set; the supply chain anomaly detection feature set includes cargo detention risk features, transportation route deviation features and order matching anomaly features.

3. The method according to claim 2, wherein The calling of the pre-trained supply chain trend prediction model, performing cross-modal feature fusion processing on the logistics node dynamic feature set and the order interaction semantic feature set, and generating the supply chain trend prediction feature set includes: Performing time alignment processing on the cargo stacking density change feature and the transportation vehicle stay time feature in the logistics node dynamic feature set to generate a first time series feature vector; Performing semantic coding processing on the demand keyword distribution features and supplier response timeliness features in the order interaction semantic feature set to generate a second time series feature vector; An attention weight allocation mechanism is used to dynamically adjust the weights of the first time series feature vector and the second time series feature vector to generate a fused cross-modal time series feature vector; wherein the dynamic weight adjustment process includes assigning different attention weight coefficients based on the degree of matching between the timestamp of the logistics node monitoring data and the timestamp of the order interaction record; The cross-modal time series feature vector is input into the supply chain trend prediction model, and the trend change patterns of different time granularities are extracted through a multi-layer time convolutional network to generate the supply chain trend prediction feature set.

4. The method according to claim 2, wherein The method of using an anomaly detection convolutional network to perform local anomaly pattern recognition processing on the dynamic feature set of the logistics node to generate the supply chain anomaly detection feature set includes: Performing light anomaly detection processing on the ambient light intensity fluctuation characteristics in the dynamic feature set of the logistics node, extracting the occurrence time and duration of the light intensity mutation event, and generating light anomaly event characteristics; Performing density gradient analysis on the cargo bulk density change characteristics to identify abnormal bulk areas where the density change rate exceeds a preset threshold, thereby generating cargo detention risk characteristics; Performing a stay pattern clustering process on the stay duration features of the transport vehicles, marking the transport vehicles that deviate from the preset stay duration as route deviation candidates, and generating transport route deviation features; The lighting anomaly event feature, the cargo detention risk feature, and the transportation route deviation feature are associated and matched with the contract clause conflict semantic feature in the order interaction semantic feature set to generate the supply chain anomaly detection feature set.

5. The method according to claim 4, wherein After generating the supply chain anomaly detection feature set, the method further includes: Inputting the cargo detention risk features and transportation route deviation features in the supply chain anomaly detection feature set into the reinforcement learning decision optimization model to trigger the anomaly feedback learning mechanism of the reinforcement learning decision optimization model; Dynamically adjust the weights of the reward and punishment functions in the abnormal feedback learning mechanism based on the historical processing records of the cargo detention risk characteristics and the severity of the current transportation route deviation characteristics; performing incremental training on the policy network of the reinforcement learning decision optimization model based on the adjusted reward and punishment function weights to generate an updated reinforcement learning decision optimization model; The updated reinforcement learning decision optimization model is applied to the subsequent processing of the supply chain anomaly detection feature set to achieve dynamic adaptive adjustment of the multi-stage decision optimization path.

6. The method according to claim 1, wherein The multi-stage decision optimization path is used to indicate the priority sequence of supply chain resource configuration and the execution order of exception handling operations. The preset reinforcement learning decision optimization model is based on which the supply chain trend prediction feature set and the supply chain anomaly detection feature set are dynamically processed to generate a multi-stage decision optimization path, including: Performing decision priority sorting on the target goods demand distribution characteristics and logistics efficiency bottleneck prediction characteristics in the supply chain trend prediction feature set to generate a resource allocation priority sequence; Performing an abnormal impact assessment on the cargo detention risk features and transportation route deviation features in the supply chain abnormality detection feature set to generate an abnormality handling operation sequence; A reinforcement learning decision optimization model is used to perform dynamic path combination optimization processing on the resource allocation priority sequence and the exception handling operation sequence to generate an initial multi-stage decision optimization path containing multiple decision sub-paths; wherein each decision sub-path corresponds to an optimized operation combination of a supply chain link; According to the preset supply chain link weight distribution table, the execution effect of each decision sub-path is simulated and predicted, and the decision sub-path combination with the best prediction effect is selected as the multi-stage decision optimization path.

7. The method according to claim 6, wherein Generating a visual graphic analysis report based on the multi-stage decision optimization path and the supply chain data set includes: Converting the resource allocation priority sequence in the multi-stage decision optimization path into a first visual decision tree structure, and marking the demand satisfaction rate and resource consumption rate of each resource allocation node in the first visual decision tree structure; Converting the exception handling operation sequence in the multi-stage decision optimization path into a second visual decision tree structure, and marking the risk reduction rate and operation execution cost of each exception handling node in the second visual decision tree structure; Performing key frame extraction processing on the image-based logistics node monitoring data in the supply chain data set to generate a logistics state change animation, and performing spatiotemporal alignment processing on the logistics state change animation with the first and second visual decision tree structures; The aligned visual decision tree structure and the logistics status change animation are interactively matched to generate the visual graphic analysis report containing the dynamic data association relationship.

8. The method according to claim 7, wherein After generating the visual graphic analysis report, the following steps are also included: Monitoring the response operation data of the enterprise management system to the visual graphic analysis report, and extracting resource configuration adjustment data and exception handling execution data from the response operation data; Performing a difference comparison process on the resource configuration adjustment data and the resource configuration priority sequence in the multi-stage decision optimization path to generate a resource configuration strategy offset feature; Performing effect evaluation processing on the exception handling execution data and the exception handling operation sequence in the multi-stage decision optimization path to generate an exception handling effect hysteresis feature; Feeding back the resource allocation strategy offset characteristics and the exception handling effect hysteresis characteristics to the reinforcement learning decision optimization model to trigger the online strategy adjustment mechanism of the reinforcement learning decision optimization model; The multi-stage decision optimization path is updated according to the output result of the online strategy adjustment mechanism, and a visual graphic optimization report containing the updated decision path is regenerated.

9. The method according to claim 1, wherein The training process of the preset reinforcement learning decision optimization model includes: Obtain historical supply chain data sets and corresponding historical decision path execution effect data; Performing feature enhancement processing on the historical supply chain data set to generate a simulated supply chain trend prediction feature set and a simulated supply chain anomaly detection feature set; Build an initial reinforcement learning decision optimization model and set a compound reward function that includes resource allocation efficiency rewards and exception handling effect rewards; The simulated supply chain trend prediction feature set and the simulated supply chain anomaly detection feature set are input into the initial reinforcement learning decision optimization model, and through multiple rounds of alternating training of strategy exploration and strategy utilization, the reinforcement learning decision optimization model that meets the preset decision effect is generated; wherein, the strategy exploration stage includes randomly attempting decision sub-paths that have not appeared in the historical decision path execution effect data, and the strategy utilization stage includes selecting the historical optimal decision sub-path for reinforcement learning based on the feedback value of the compound reward function.

10. A digital enterprise management data analysis system, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is enabled to perform the steps of any one of the methods of claims 1 to 9.

Images