A Smart Sales System for Electronic Components Based on Demand Forecasting and Dynamic Pricing

Abstract

This invention discloses an intelligent sales system for electronic components based on demand forecasting and dynamic pricing, belonging to the field of distributed data processing technology. It includes a data acquisition module that collects microservice call chain data through a distributed link tracing system; a semantic compression module that receives high-dimensional tensors output by an external prediction model and compresses them using an autoencoder neural network to generate low-dimensional feature vectors that the pricing engine can directly consume; an early warning module that predicts the probability of future resource conflicts based on historical resource competition index sequences using a time-series prediction model; a scheduling module that constructs a coupling degree matrix of system variables and calculates coupling entropy, and executes resource scheduling strategies according to nonlinear scheduling decision factors; and a feedback optimization module that dynamically updates the early warning threshold by collecting performance data after scheduling. This invention significantly reduces system response latency and resource consumption.

Description

Technical Field

[0001] This invention belongs to the field of distributed data processing technology, and in particular relates to an intelligent sales system for electronic components based on demand forecasting and dynamic pricing. Background Technology

[0002] In existing technologies, intelligent sales systems under a microservice architecture typically employ distributed tracing technology to monitor service call status and optimize system performance through resource scheduling strategies. For example, traffic-aware adaptive scheduling frameworks dynamically adjust service deployment locations to optimize network transmission overhead by monitoring inter-service communication latency and load pressure in real time. Furthermore, adaptive management methods based on the MAPE-K closed-loop control architecture are also widely used in microservice systems to maintain system performance stability by monitoring service quality indicators and performing corresponding scaling up / down or rescheduling operations.

[0003] However, existing technologies still have the following drawbacks: First, there is a semantic gap between the high-dimensional tensor data output by the prediction model and the input format required by the pricing engine, resulting in continuous serialization and deserialization overhead at the interface, which cannot be eliminated even with optimization of the scheduling strategy; Second, existing scheduling strategies mostly adopt a passive response mechanism, that is, scheduling actions are triggered only after performance degradation is detected, which cannot preventive intervention before resource conflicts occur; In addition, multi-index scheduling decisions usually use linear weighting or simple threshold rules, which are difficult to accurately capture the nonlinear coupling relationship between system variables, resulting in limited accuracy of scheduling decisions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent sales system for electronic components based on demand forecasting and dynamic pricing, thus solving the aforementioned problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent sales system for electronic components based on demand forecasting and dynamic pricing, comprising:

[0006] The data acquisition module collects microservice call chain data through a distributed tracing system. The call chain data includes service node processing time and cross-service call data transmission volume.

[0007] The semantic compression module receives a high-dimensional tensor output by an external prediction model and performs feature compression on the high-dimensional tensor through an autoencoder neural network to generate a low-dimensional feature vector that the pricing engine can directly consume, thereby eliminating runtime format conversion overhead.

[0008] The early warning module, based on the historical resource competition index sequence, uses a time-series prediction model to predict the probability of resource conflicts within future time windows;

[0009] The scheduling module constructs the coupling degree matrix of system variables, obtains the coupling entropy, and executes resource scheduling strategies based on nonlinear scheduling decision factors.

[0010] The feedback optimization module collects performance data after the scheduling strategy is executed and dynamically updates the warning threshold and scheduling parameters.

[0011] The autoencoder includes an encoder and a decoder. The encoder consists of multiple fully connected layers and nonlinear activation functions, used to map high-dimensional inputs to low-dimensional latent spaces. The decoder consists of multiple fully connected layers and nonlinear activation functions, used to reconstruct low-dimensional latent variables into the original high-dimensional space.

[0012] Based on the above technical solutions, the present invention also provides the following optional technical solutions:

[0013] Further technical solution: The loss function used by the semantic compression module is:

[0014]

[0015] in, This indicates the loss function used by the semantic compression module. This represents a high-dimensional tensor received from an external prediction model. The encoder function (consisting of multiple fully connected layers and nonlinear activation functions, which maps the input to a low-dimensional latent space) represents the encoder function. The decoder function (consisting of multiple fully connected layers and non-linear activation functions, which reconstructs the latent variables into their original dimensions) represents the decoder function. This represents the squared L2 norm (used to measure reconstruction error). This represents the L1 norm (used to constrain the sparsity of hidden layer features). This represents the sparse regularization coefficient.

[0016] Further technical solution: The predictive early warning module uses a Long Short-Term Memory (LSTM) network as the time-series prediction model, and inputs historical data... The resource competition index sequence for each time window outputs the future... The predicted value for each time window; the formula for calculating the probability of resource conflict is:

[0017]

[0018] in, This indicates the probability of resource conflicts occurring within the future window. This represents the future output of the LSTM model. The predicted value of the resource competition index for each time window. This represents the preset resource competition warning threshold. This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0019] Further technical solution: The coupling matrix constructed by the scheduling module is as follows:

[0020]

[0021] in, Represents the coupling degree matrix. The Pearson correlation coefficient represents the relationship between equivalent processing time and the resource competition index. The Pearson correlation coefficient represents the relationship between equivalent processing time and cache benefit factor. The Pearson correlation coefficient represents the relationship between the resource contention index and the cache benefit coefficient. Each Pearson correlation coefficient is calculated based on the sampling data of the most recent time window, and the coupling matrix is ​​a symmetric matrix.

[0022] A further technical solution: The formula for obtaining the coupling entropy in the scheduling module is:

[0023]

[0024] in, Represents coupling entropy, This represents an element in the coupling degree matrix.

[0025] A further technical solution: The formula for obtaining the nonlinear scheduling decision factor in the scheduling module is as follows:

[0026]

[0027] in, Represents the nonlinear scheduling decision factor. Represents coupling entropy, Maximum allowable coupling entropy This represents the resource competition index. This represents the maximum permissible resource competition index. Indicates the equivalent processing time. Indicates the equivalent processing time threshold. This represents a nonlinear exponent.

[0028] A further technical solution: The formula for obtaining the resource competition index is as follows:

[0029]

[0030] in, This represents the resource competition index. This represents the number of instructions per clock cycle for the predictive model service process. This represents the number of instructions per clock cycle for the pricing engine service process. This represents the system's theoretical maximum number of instructions per clock cycle. This represents the number of access conflicts in the shared LLC cache per unit of time. This represents the total number of accesses to the shared LLC cache per unit of time.

[0031] Further technical solution: The cache benefit coefficient is obtained by using the following formula based on the cache hit rate, the current system throughput, and the cache memory size:

[0032]

[0033] in, This represents the caching benefit coefficient. Indicates cache hit rate. This indicates the current system throughput. This indicates the memory size occupied by the cache. Indicates the equivalent processing time.

[0034] A further technical solution: The equivalent processing time is obtained by: based on the semantically compressed first... The weighted fan-out coefficient is obtained by using the following formula, based on the amount of data transferred in each call and the preset baseline data amount:

[0035]

[0036] in, This represents the weighted fan-out coefficient. Indicates the number of cross-service calls. Indicates the semantically compressed first... Data transfer volume per call This indicates the preset baseline data volume.

[0037] Further technical solutions: Based on the first The equivalent processing time is obtained by considering the processing time of each service node, the available bandwidth of the root node, and the weighted fan-out coefficient, using the following formula:

[0038]

[0039] in, Indicates the equivalent processing time. Indicates the total number of service nodes. Indicates the number of cross-service calls. Indicates the first Processing time of each service node Available bandwidth for the root node Indicates the semantically compressed first... Data transfer volume per call Indicates the network protocol overhead factor. This represents the weighted fan-out coefficient.

[0040] Further technical solution: The rule for the feedback optimization module to update the resource contention early warning threshold is as follows:

[0041]

[0042] in, This indicates the updated resource contention warning threshold. This represents the rate of change of the equivalent processing time before and after scheduling. Indicates the threshold adjustment step size. This indicates the resource contention warning threshold before the update;

[0043] The rate of change of the equivalent processing time before and after scheduling is expressed as:

[0044]

[0045] in, This represents the rate of change of the equivalent processing time before and after scheduling. This represents the equivalent processing time before scheduling. This indicates the equivalent processing time after scheduling.

[0046] This invention provides an intelligent sales system for electronic components based on demand forecasting and dynamic pricing, which has the following advantages compared with the prior art:

[0047] This invention encodes the high-dimensional tensor output by the prediction model into a low-dimensional feature vector that the pricing engine can directly consume through a semantic compression module. This fundamentally eliminates the runtime serialization and deserialization overhead, significantly reduces the amount of cross-service data transmission, substantially shortens the equivalent processing time, and effectively improves the system response speed.

[0048] This invention uses a predictive early warning module to predict the probability of future resource conflicts based on historical resource competition index sequences and a time-series prediction model. This upgrades the scheduling mode from the traditional passive response mode to an active prevention mode. It can identify potential risks in time before resource conflicts occur and execute preventive scheduling, effectively avoiding performance fluctuations during peak business periods and enhancing system stability.

[0049] This invention constructs a coupling degree matrix containing equivalent processing time, resource contention index, and cache benefit coefficient through a nonlinear scheduling module, calculates coupling entropy, and generates nonlinear scheduling decision factors. This can accurately capture the nonlinear coupling relationship between system variables, avoid the decision bias caused by the traditional linear weighting method, and significantly improve the accuracy and robustness of scheduling decisions.

[0050] This invention uses a feedback optimization module to dynamically update the early warning threshold based on the rate of performance change before and after scheduling, forming a closed-loop adaptive adjustment mechanism. This enables the system to automatically adapt to dynamic changes in business load and maintain high-performance and stable operation for a long time without manual intervention, thereby reducing operation and maintenance costs.

[0051] This invention effectively solves the technical problem of low data interaction efficiency between the prediction model and the pricing engine, significantly reduces system resource consumption, greatly improves data processing real-time performance and system throughput, and enhances user experience and transaction conversion efficiency. Attached Figure Description

[0052] Figure 1 This is the intended flow of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0054] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0055] Please see Figure 1 According to one embodiment of the present invention, an intelligent sales system for electronic components based on demand forecasting and dynamic pricing includes:

[0056] The data acquisition module collects microservice call chain data through a distributed tracing system. The call chain data includes service node processing time and cross-service call data transmission volume.

[0057] The semantic compression module receives a high-dimensional tensor output by an external prediction model and performs feature compression on the high-dimensional tensor through an autoencoder neural network to generate a low-dimensional feature vector that the pricing engine can directly consume, thereby eliminating runtime format conversion overhead.

[0058] The early warning module, based on the historical resource competition index sequence, uses a time-series prediction model to predict the probability of resource conflicts within future time windows;

[0059] The scheduling module constructs the coupling degree matrix of system variables, obtains the coupling entropy, and executes resource scheduling strategies based on nonlinear scheduling decision factors.

[0060] The feedback optimization module collects performance data after the scheduling strategy is executed and dynamically updates the warning threshold and scheduling parameters.

[0061] The autoencoder includes an encoder and a decoder. The encoder consists of multiple fully connected layers and nonlinear activation functions, used to map high-dimensional inputs to low-dimensional latent spaces. The decoder consists of multiple fully connected layers and nonlinear activation functions, used to reconstruct low-dimensional latent variables into the original high-dimensional space.

[0062] Specifically, the system includes a data acquisition module configured to collect microservice call chain data through a distributed tracing system. For example, the data acquisition module can be configured to periodically poll the log files of each microservice instance to extract the call relationships and timestamp information between services, or to record the start time, end time, and data packet size of requests by inserting probes at each service entry and exit point. The call chain data further includes service node processing time and cross-service call data transmission volume. Service node processing time can be obtained simply by recording the time difference between the request entering and leaving the service node, while cross-service call data transmission volume can be obtained by monitoring network interface traffic statistics or by counting bytes at the data transmission layer.

[0063] The system also includes a semantic compression module configured to receive high-dimensional tensors output by external predictive models. For example, an external predictive model might output a vector containing thousands of features describing information such as market trends, inventory levels, and customer behavior. This semantic compression module can compress these high-dimensional tensors using traditional dimensionality reduction algorithms such as Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA). These methods project high-dimensional data into a low-dimensional space, generating low-dimensional feature vectors that the pricing engine can directly consume. This process aims to eliminate runtime format conversion overhead, as the pricing engine can directly process these pre-processed, uniformly formatted low-dimensional vectors without complex serialization and deserialization operations.

[0064] In addition, the system includes an early warning module, which is configured to predict the probability of resource conflict within a future time window based on a historical resource competition index sequence using a time series forecasting model. For example, the historical resource competition index sequence can be composed of indicators such as CPU utilization, memory usage, or network latency in the past hour. The early warning module can use time series forecasting models such as the autoregressive moving average model (ARMA) or exponential smoothing to analyze these historical sequences, thereby predicting the trend of the resource competition index in the next few hours. Based on the prediction results, the probability of resource conflict occurring in a certain period of time in the future can be calculated.

[0065] The system further includes a scheduling module configured to construct a coupling matrix of system variables. This coupling matrix can be constructed based on key performance indicators in the system, such as response time, throughput, and error rate, by calculating their correlation coefficients. The scheduling module is also configured to obtain coupling entropy, which can be obtained by performing mathematical operations on the elements of the coupling matrix. For example, the absolute values ​​of all non-zero elements in the matrix can be simply added to reflect the overall correlation of the system variables. Subsequently, the scheduling module executes a resource scheduling strategy based on a nonlinear scheduling decision factor. This nonlinear scheduling decision factor can be a weighted sum based on the current system load, resource utilization, and warning status. For example, when the system load is high, this factor may prompt the scheduling module to prioritize expansion operations; when resource utilization is low, it may trigger shrinkage operations.

[0066] Finally, the system includes a feedback optimization module, which is configured to collect performance data after the scheduling policy is executed. For example, after the scheduling policy is executed, this module can collect new data such as service response time, resource utilization, and request processing volume. Based on this performance data, the feedback optimization module is configured to dynamically update the warning threshold and scheduling parameters. For example, if the system performance is significantly improved after the scheduling policy is executed, the warning threshold can be appropriately relaxed; conversely, if the performance deteriorates or resource conflicts are frequent, the warning threshold may be tightened. The updating of scheduling parameters can be based on simple heuristic rules. For example, if a certain scheduling parameter causes a performance bottleneck, it can be fine-tuned in the next scheduling.

[0067] The following example will provide a more detailed explanation of the above technical solution:

[0068] Imagine a smart e-commerce platform for electronic components, where user A is browsing and purchasing products. The platform employs a microservice architecture, comprising multiple microservices such as product recommendation, inventory management, order processing, and pricing engine. To ensure a smooth shopping experience for user A, the system needs to efficiently handle a large number of concurrent requests and dynamically adjust product prices based on real-time demand.

[0069] First, the data acquisition module works continuously, collecting microservice call chain data through a distributed tracing system. For example, when user A initiates a product query request, the request will sequentially pass through the front-end service, product recommendation service, inventory management service, and finally reach the pricing engine service. The data acquisition module records the time spent by each service node in processing the request, as well as the amount of data such as product information and inventory data transmitted between services. This data is collected in real time, forming detailed microservice call chain data, providing a foundation for subsequent analysis.

[0070] Simultaneously, the external prediction model outputs prediction results in the form of a high-dimensional tensor based on information such as market supply and demand and historical sales data. For example, it predicts the future demand and price fluctuation trend of a certain electronic component. This high-dimensional tensor is sent to the semantic compression module. Upon receiving the high-dimensional tensor, the semantic compression module immediately performs feature compression on it using an autoencoder neural network. For instance, a high-dimensional tensor containing hundreds of market indicators is transformed into a low-dimensional latent variable containing only a few dozen key features after processing by the encoder. Subsequently, the decoder attempts to reconstruct the original high-dimensional tensor from this low-dimensional latent variable, optimizing the encoder's compression effect by comparing the reconstruction error. Finally, the semantic compression module outputs a low-dimensional feature vector, which directly contains the core information required by the pricing engine, such as the predicted demand level and suggested price range for the component. This low-dimensional feature vector is directly passed to the pricing engine, thus avoiding the serialization and deserialization overhead caused by the mismatch between high-dimensional data and the pricing engine's input format in traditional methods.

[0071] During this process, the early warning module also runs continuously. Based on historical resource contention index sequences, such as the average CPU utilization, memory usage, and network latency over the past 24 hours, this module uses a time-series prediction model to predict the probability of resource conflicts within future time windows. For example, if the prediction model shows that there is an 80% probability that CPU utilization will exceed a preset threshold within the next hour due to a surge in requests caused by promotional activities, the early warning module will issue an alert for potential resource conflicts.

[0072] When the early warning module issues an alert or the system load changes significantly, the scheduling module is activated. This module first constructs a coupling matrix of system variables; for example, this matrix might reflect the degree of interaction between three key system variables: CPU utilization, memory usage, and network bandwidth. Next, the scheduling module obtains the coupling entropy, which quantifies the overall complexity of the correlation between these variables. Based on this coupling entropy and other system state information, the scheduling module calculates a nonlinear scheduling decision factor. For example, if the coupling entropy is high and the resource contention index is close to a threshold, this factor indicates that the system is currently in a high-risk state and a more aggressive scheduling strategy is needed. According to this nonlinear scheduling decision factor, the scheduling module executes a resource scheduling strategy. For example, if CPU resource conflict is predicted to occur, the scheduling module might decide to increase the number of instances of the product recommendation service or migrate some low-priority tasks to idle servers.

[0073] After the scheduling strategy is executed, the feedback optimization module immediately begins collecting performance data. For example, this module monitors whether the average response time of the product recommendation service decreases after expansion, and whether the overall system throughput increases. Based on this real-time performance data, the feedback optimization module dynamically updates the warning thresholds and scheduling parameters. For instance, if system performance significantly improves after expansion and the probability of resource conflicts decreases, the feedback optimization module may appropriately raise the warning threshold for CPU utilization to allow the system to run under higher loads. Simultaneously, the triggering conditions for expansion in the scheduling parameters may also be adjusted to make them more sensitive or more conservative to adapt to the new system operating conditions. Through this closed-loop feedback mechanism, the system can continuously learn and optimize its operating strategies, thereby achieving intelligent sales and resource management.

[0074] First, addressing the semantic gap between the high-dimensional tensor output by the prediction model and the input format of the pricing engine in existing technologies, and the resulting runtime format conversion overhead, this embodiment introduces a semantic compression module and utilizes an autoencoder neural network to compress the features of the high-dimensional tensor, generating a low-dimensional feature vector that the pricing engine can directly consume. In the example above, the complex market prediction data output by the external prediction model is efficiently converted into a simplified format that the pricing engine can directly process, effectively avoiding the frequent serialization and deserialization operations in traditional solutions. Compared to the inherent overhead that cannot be eliminated even with optimized scheduling strategies in existing technologies, this embodiment solves the format mismatch problem at the data source, significantly improving data processing efficiency and system response speed.

[0075] Secondly, addressing the shortcomings of existing scheduling strategies that mostly employ passive response mechanisms and cannot preventatively intervene before resource conflicts occur, this embodiment designs an early warning module. This module uses a time-series prediction model based on historical resource contention index sequences to predict the probability of resource conflicts within future time windows. In the example, the system can predict CPU resource conflicts that may arise due to promotional activities and issue timely alerts. This contrasts sharply with the lagging mechanism in existing technologies that only triggers scheduling actions after performance degradation is detected, enabling the system to intervene preventatively and thus avoiding potential performance bottlenecks and a decline in user experience.

[0076] Furthermore, addressing the issue that existing technologies often employ linear weighting or simple threshold rules for multi-indicator scheduling decisions, which struggle to accurately capture the nonlinear coupling relationships between system variables and thus limit the accuracy of scheduling decisions, this embodiment's scheduling module constructs a coupling degree matrix of system variables, obtains coupling entropy, and executes resource scheduling strategies based on nonlinear scheduling decision factors. In this example, the scheduling module comprehensively considers the complex relationships between multiple variables such as CPU utilization, memory usage, and network bandwidth, and makes more refined scheduling decisions based on nonlinear factors. For instance, it may take more aggressive expansion measures under specific high-risk conditions. This surpasses the simple linear weighting or threshold judgments of existing technologies, enabling scheduling decisions to more accurately reflect the actual complex dynamics of the system and improving the effectiveness and accuracy of resource scheduling.

[0077] Finally, the feedback optimization module in this embodiment dynamically updates the warning threshold and scheduling parameters by collecting performance data after the scheduling strategy is executed, thus achieving adaptive optimization of the system. In the example, the system can intelligently adjust the warning threshold for CPU utilization and the expansion trigger conditions based on the actual performance after expansion. This closed-loop feedback mechanism enables the system to continuously learn and adapt to the constantly changing operating environment, continuously optimizing its performance, thereby ensuring the long-term stable and efficient operation of the intelligent sales system. This deep adaptive capability is lacking in existing technologies.

[0078] In summary, this embodiment constructs an intelligent sales system for electronic components that can proactively predict, make intelligent decisions, and continuously optimize through the collaborative work of multiple modules such as data acquisition, semantic compression, prediction and early warning, nonlinear scheduling decision-making, and feedback optimization. It effectively solves the problems of semantic gap, passive response, and insufficient decision-making accuracy in the prior art, demonstrating significant technological progress and practical value.

[0079] Preferably, the loss function used by the semantic compression module is:

[0080]

[0081] in, This indicates the loss function used by the semantic compression module. This represents a high-dimensional tensor received from an external prediction model. The encoder function (consisting of multiple fully connected layers and nonlinear activation functions, which maps the input to a low-dimensional latent space) represents the encoder function. The decoder function (consisting of multiple fully connected layers and non-linear activation functions, which reconstructs the latent variables into their original dimensions) represents the decoder function. This represents the squared L2 norm (used to measure reconstruction error). represents the L1 norm (used to constrain the sparsity of hidden layer features), and represents the sparsity regularization coefficient.

[0082] The nonlinear activation function used in the encoder and decoder described in this invention can be any nonlinear activation function known in the art, including but not limited to ReLU (Rectified Linear Unit), Leaky ReLU, Sigmoid, Tanh, etc. Preferably, the hidden layer uses the ReLU activation function, which is expressed as ReLU(x) = max(0,x), because it is simple to calculate and can effectively alleviate the gradient vanishing problem.

[0083] The loss function is a core component of machine learning model training. Its role is to quantify the difference between the model's predicted output and the true target, guiding the model to adjust its internal parameters to minimize reconstruction error and achieve specific feature representation goals. Besides the form combining L2 reconstruction error and L1 sparse regularization used in this invention, the loss function can also be mean squared error, cross-entropy, or combined with other regularization terms such as L2 regularization. High-dimensional tensors It is a multidimensional array used to represent complex data from an external prediction model, with the aim of compressing it into a low-dimensional space. Encoder function. Responsible for inputting high-dimensional tensors The input data is mapped to a low-dimensional latent space, achieved through a series of fully connected layers and nonlinear activation functions, thereby extracting the core features. The encoder function can consist of a multilayer perceptron, a convolutional neural network, or a recurrent neural network, etc. The decoder function... The purpose is to reconstruct the original high-dimensional space from the low-dimensional latent variables generated by the encoder. This is achieved through a series of fully connected layers and nonlinear activation functions, aiming to recover the original input information as much as possible. (L2 norm squared) In loss functions, the L2 norm squared loss is commonly used to measure reconstruction error. It calculates the sum of squared differences between corresponding elements of the reconstructed output and the original input, penalizing larger errors more severely. Besides the L2 norm squared loss, the L1 norm or Huber loss can also be used to measure reconstruction error. In loss functions, sparsity regularization is often used to constrain the sparsity of hidden layer features. By minimizing the L1 norm, the model tends to push most feature values ​​in the hidden layer towards zero, thus activating only a few important features, achieving feature selection and dimensionality reduction. Besides the L1 norm, sparsity constraints can also be achieved through methods such as KL divergence. (Sparse regularization coefficient) It is a hyperparameter used to balance the importance of the reconstruction error term and the sparsity constraint term in the loss function, and is usually tuned by methods such as cross-validation or grid search.

[0084] The solution of this invention introduces a specific loss function. To optimize the autoencoder neural network in the semantic compression module, the loss function consists of two parts: the first part is the reconstruction error term. It measures how an autoencoder renders a high-dimensional tensor. After being encoded as low-dimensional latent variables and reconstructed by a decoder, the difference between the reconstructed result and the original input is minimized. This allows the autoencoder to be trained to retain information from the original data as accurately as possible. The second part is the sparsity constraint term. It applies an L1 norm penalty to the low-dimensional latent variables of the encoder output. This term encourages the autoencoder to learn sparse feature representations, meaning that in the low-dimensional latent space, most feature values ​​approach zero, and only a few key features are activated. The sparse regularization coefficient... The weights of these two parts of the loss are then adjusted to ensure effective feature sparsity while maintaining reconstruction accuracy. In this way, the autoencoder can generate low-dimensional feature vectors that accurately reflect the original high-dimensional information and have good sparsity, thus providing efficient and easy-to-process input for the pricing engine, effectively reducing runtime format conversion overhead, and improving the overall system efficiency and robustness.

[0085] As a specific implementation method, the semantic compression module receives the high-dimensional tensor output by the external prediction model. Then, it will be input into a pre-trained autoencoder. The training process of the autoencoder is as follows: First, the high-dimensional tensor... After encoder function Processing to generate low-dimensional latent variables Subsequently, this low-dimensional latent variable is passed through the decoder function. Process and reconstruct it into a tensor in the original high-dimensional space. In each training iteration, the system calculates the loss function for the current batch of data. Specifically, the reconstruction error term By calculating the original tensor With the reconstruction tensor It is obtained by summing the squared differences between corresponding elements. Simultaneously, the sparsity constraint term... By calculating low-dimensional latent variables The sum of the absolute values ​​of all elements in the expression, multiplied by a preset sparsity regularization coefficient. To obtain. The sum of these two items gives the total loss. Then, using the backpropagation algorithm and an optimizer (such as the Adam optimizer or SGD optimizer), based on the loss function... The gradient is used to update the weight parameters in the encoder and decoder. This training process continues until the loss function converges or a preset number of training epochs are reached. Through this training method, the autoencoder can learn an efficient feature compression strategy, so that the generated low-dimensional feature vectors retain key information while also being sparse.

[0086] Through the above technical solution, the semantic compression module can compress the features of the high-dimensional tensor output by the external prediction model using a carefully designed loss function. This effectively balances reconstruction accuracy and feature sparsity. The loss function combines the squared L2 norm to accurately measure reconstruction error and introduces the L1 norm to impose sparsity constraints on hidden layer features, supplemented by a sparsity regularization coefficient. This adjustment allows the autoencoder to learn an optimized feature representation, where the generated low-dimensional feature vectors not only highly reproduce the original high-dimensional information but also exhibit significant sparsity. This sparsity helps reduce feature redundancy, improves the model's generalization ability, and reduces the computational complexity and storage requirements of the subsequent pricing engine when processing these features. Ultimately, this further enhances the semantic compression module's ability to eliminate runtime format conversion overhead, providing the pricing engine with more efficient and concise input, thereby improving the response speed and resource utilization efficiency of the entire intelligent electronic component sales system.

[0087] Preferably, the predictive early warning module uses a Long Short-Term Memory (LSTM) network as the time-series prediction model, and inputs historical data... The resource competition index sequence for each time window outputs the future... The predicted value for each time window; the formula for calculating the probability of resource conflict is:

[0088]

[0089] in, This indicates the probability of resource conflicts occurring within the future window. This represents the future output of the LSTM model. The predicted value of the resource competition index for each time window. This represents the preset resource competition warning threshold. This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0090] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network. Their core lies in using gating mechanisms (including input gates, forget gates, and output gates) to control the flow and retention of information within the network, effectively solving the gradient vanishing or exploding problems faced by traditional recurrent neural networks when processing long sequences of data. In this invention, LSTM is used as a time-series prediction model, capable of deeply learning the complex time dependencies and nonlinear patterns inherent in resource competition index sequences, thereby improving the accuracy and robustness of predictions. This LSTM model can be constructed as a multi-layer structure, with each layer containing multiple LSTM units to capture temporal features at different levels of abstraction; alternatively, a bidirectional LSTM structure can be used, simultaneously utilizing forward and backward information from the sequence to enhance the understanding of changes in the resource competition index. Input History A resource competition index sequence within a time window refers to a sequence of resource competition indices over a past period (e.g., recent timeframes). Resource competition index data monitored over several sampling periods are used as input to the LSTM model. This historical data forms the basis for model learning and prediction, and can be continuously acquired from the data acquisition module and preprocessed. For example, the resource competition index can be recorded at fixed time intervals (e.g., 10 minutes or 1 hour), and the most recent data can be selected. The input sequence can be composed of consecutive data points; alternatively, a sliding window mechanism can be used to dynamically extract data from a longer historical dataset. Data from a given time window is used as input. Output the future... The predicted values ​​for each time window are the expected estimates of the resource competition index over a future period based on the learned patterns of the LSTM model. These predictions are crucial for subsequent calculations of the probability of resource conflict. LSTM models can be designed to directly output future... Alternatively, a multi-step prediction method can be used, where the model predicts one time step at a time and uses that prediction as input for the next prediction, repeating this process. This will allow us to obtain the complete future. A predicted sequence for each time window. Formula for calculating the probability of resource conflicts. This provides an effective method for quantifying the risk of future resource conflicts. The formula uses statistical analysis of future... Within a time window, the predicted resource competition index Exceeding the preset resource contention warning threshold The frequency of the index is calculated and averaged to intuitively reflect the likelihood of the system facing resource pressure in the future. Indicator function It is a binary function, when its internal conditions (i.e. When the condition is true, the output value is 1, indicating that there is a risk of resource conflict within that time window; otherwise, the output value is 0. In actual implementation, the indicator function can be implemented through conditional statements (such as if-else structures) or mapping Boolean values ​​to numerical values.

[0091] This invention effectively handles the complex and long-term dependent historical resource competition index sequences generated during the operation of an intelligent sales system for electronic components by introducing a Long Short-Term Memory (LSTM) network as the time-series prediction model for the early warning module. The microservice call chain data continuously collected by the data acquisition module is processed to form a historical resource competition index sequence, which serves as the input to the LSTM model and includes the most recent... Resource competition index data for each time window. The LSTM model utilizes its unique gating mechanism to perform deep learning and pattern recognition on this historical data, thereby accurately predicting the future. Resource competition index within a time window Once the future is obtained Based on the predicted resource competition index for each time window, the early warning module uses a specific calculation formula to quantify the probability of future resource conflicts. The calculation formula works by traversing the future... The system divides the time window into several time windows and forecasts the resource competition index within each time window. Compared with the preset resource competition warning threshold Compare the predicted values. Exceeding the threshold At that time, indicator function Output 1 if there is a potential resource conflict within the time window; otherwise, output 0. Finally, process all future... The number of times the indicator function outputs 1 within each time window is accumulated and then divided by 1. This allows us to obtain the average probability of resource conflicts occurring within the future window. This approach, combining LSTM model prediction with probabilistic formulas for quantification, enables the early warning module to more accurately and timely assess future resource pressures on the system. Compared to using only general time-series prediction models, LSTM is better suited to the complex and variable workload characteristics of intelligent electronic component sales systems, thus providing more reliable early warning information. This precise early warning information can then be utilized by the scheduling and feedback optimization modules to guide the formulation and adjustment of dynamic pricing and resource scheduling strategies, thereby avoiding potential system performance bottlenecks and resource contention issues and ensuring the stable and efficient operation of the system.

[0092] The following is a concrete example to illustrate this. The early warning module can be configured with a network containing two layers of LSTM units, each with 100 and 50 neurons respectively, supplemented by a Dropout layer to prevent overfitting, and finally connected to a fully connected layer to output the predicted value. This LSTM model can be trained using a resource competition index sequence from the past 7 days (e.g., sampled hourly, for a total of 168 data points) to learn its periodic changes and sudden fluctuations. In actual operation, the system collects the latest resource competition index hourly and compares it with the previous... The data from 23 hours are combined to form a historical sequence of length 24, which is then used as input to the LSTM model. Assume the system needs to predict the future. = 4-hour resource competition index. The LSTM model will output predicted resource competition index values ​​for the next 4 time points based on this 24-hour historical data, for example... =0.65, =0.82, =0.78, =0.70. Meanwhile, the system's preset resource contention warning threshold... The value is 0.75. The early warning module will then check each of these four predicted values. =0.65, because 0.65 0.75, indicator function Output 0. For =0.82, since 0.82 > 0.75, the indicator function Output 1. For =0.78$, since 0.78 > 0.75, the indicator function Output 1. For =0.70, because 0.70 0.75, indicator function Output 0. Ultimately, the resource conflict probability... The calculation is (0+1+1+0) / 4 = 2 / 4 = 0.5. This calculation result indicates that there is a 50% probability that the system will encounter resource conflicts within the next 4 hours, thus providing a quantitative basis for the scheduling module to take intervention measures in advance.

[0093] Through the aforementioned technical solution, the early warning module employs a Long Short-Term Memory (LSTM) network as its time-series prediction model. This effectively captures the complex nonlinear characteristics and long-term dependencies in the resource competition index sequence of the intelligent sales system for electronic components, significantly improving the accuracy of future resource competition index predictions. Based on this, and combined with a clearly defined formula for calculating the probability of resource conflicts, the system can quantitatively compare the future predicted values ​​output by the LSTM model with preset early warning thresholds, thereby accurately assessing the risk of resource conflicts occurring within future time windows. This precise and timely early warning capability overcomes the limitations of traditional time-series prediction models when processing complex system data, enabling the system to identify potential resource bottlenecks earlier and more accurately. This provides a solid data foundation for the scheduling module to formulate forward-looking resource scheduling strategies and dynamic pricing decisions, helping the system maintain high performance and stability during demand fluctuations, effectively avoiding performance degradation or service interruptions caused by resource contention, and thus optimizing the overall operational efficiency and user experience of the intelligent sales system.

[0094] Preferably, the coupling matrix constructed by the scheduling module is:

[0095]

[0096] in, Represents the coupling degree matrix. The Pearson correlation coefficient represents the relationship between equivalent processing time and the resource competition index. The Pearson correlation coefficient represents the relationship between equivalent processing time and cache benefit factor. The Pearson correlation coefficient represents the relationship between the resource contention index and the cache benefit coefficient. Each of the Pearson correlation coefficients is calculated based on the sampled data of the most recent time window, and the coupling matrix is ​​a symmetric matrix.

[0097] The formula for obtaining the coupling entropy in the scheduling module is:

[0098]

[0099] in, Represents coupling entropy, Represents the elements in the coupling degree matrix;

[0100] The formula for obtaining the nonlinear scheduling decision factor in the scheduling module is as follows:

[0101]

[0102] in, Represents the nonlinear scheduling decision factor. Represents coupling entropy, Maximum allowable coupling entropy This represents the resource competition index. This represents the maximum permissible resource competition index. Indicates the equivalent processing time. Indicates the equivalent processing time threshold. Indicates a nonlinear exponent;

[0103] The formula for obtaining the resource competition index is:

[0104]

[0105] in, This represents the resource competition index. This represents the number of instructions per clock cycle for the predictive model service process. This represents the number of instructions per clock cycle for the pricing engine service process. This represents the system's theoretical maximum number of instructions per clock cycle. This represents the number of access conflicts in the shared LLC cache per unit of time. This represents the total number of accesses to the shared LLC cache per unit of time.

[0106] The cache benefit coefficient is obtained as follows: based on the cache hit rate, the current system throughput, and the cache memory size, the cache benefit coefficient is obtained using the following formula:

[0107]

[0108] in, This represents the caching benefit coefficient. Indicates cache hit rate. This indicates the current system throughput. This indicates the memory size occupied by the cache. Indicates the equivalent processing time;

[0109] The equivalent processing time is obtained by: based on the semantically compressed first... The weighted fan-out coefficient is obtained by using the following formula, based on the amount of data transferred in each call and the preset baseline data amount:

[0110]

[0111] in, This represents the weighted fan-out coefficient. Indicates the number of cross-service calls. Indicates the semantically compressed first... Data transfer volume per call Indicates the preset baseline data volume;

[0112] Based on the The equivalent processing time is obtained by considering the processing time of each service node, the available bandwidth of the root node, and the weighted fan-out coefficient, using the following formula:

[0113]

[0114] in, Indicates the equivalent processing time. Indicates the total number of service nodes. Indicates the number of cross-service calls. Indicates the first Processing time of each service node Available bandwidth for the root node Indicates the semantically compressed first... Data transfer volume per call Indicates the network protocol overhead factor. This represents the weighted fan-out coefficient.

[0115] The coupling degree matrix is ​​a mathematical tool used to quantify the degree of interdependence between different variables in a system. The elements of the matrix represent the correlation or influence strength between variable pairs. This matrix can be calculated using statistical methods, such as Pearson correlation coefficient and Spearman correlation coefficient, to determine the linear or nonlinear correlations between different system variables (e.g., equivalent processing time, resource competition index, cache benefit coefficient), and these correlation coefficients can be used as elements of the matrix. Alternatively, machine learning models, such as decision trees, random forests, or neural networks, can be used to learn and quantify complex dependencies between variables, thereby constructing or filling the coupling degree matrix to reveal the mutual influence of variables within the system and provide a quantitative basis for subsequent scheduling decisions. Coupling entropy is an indicator that measures the degree of coupling between system variables. A higher value generally indicates stronger interdependence between parts of the system, more complex system behavior, and greater difficulty in prediction and control. This coupling entropy can be calculated using specific mathematical formulas based on the elements of the coupling degree matrix, as described in the above scheme. Alternatively, other information theory indicators, such as mutual information and conditional entropy, can be used to quantify the amount of shared information between variables, thereby indirectly reflecting the degree of coupling and providing a single numerical value to summarize the overall coupling complexity of the system, serving as an important input for scheduling decisions. The nonlinear scheduling decision factor is a comprehensive indicator used to guide the scheduling module in making dynamic resource scheduling decisions in complex system environments. Its "nonlinear" characteristic means that the factor can more precisely reflect subtle changes in the system state and adjust the scheduling strategy accordingly, rather than a simple linear mapping. This factor can be calculated by combining multiple system state parameters such as coupling entropy, resource competition index, and equivalent processing time, and by introducing a nonlinear index for weighted calculation. In addition, it can also be dynamically generated based on the current system state and preset scheduling objectives through methods such as reinforcement learning and fuzzy logic systems. This integrates multiple key system indicators into a unified decision basis, enabling the scheduling strategy to respond to system changes more flexibly and intelligently. The resource contention index is an indicator that measures the intensity of competition for shared resources (such as CPU and cache) among different service processes within a system. A high contention index usually indicates the risk of resource bottlenecks or performance degradation. This index can be calculated by comparing the differences in the number of instructions per clock cycle among different service processes and combining the ratio of the number of access conflicts to the total number of accesses to the shared cache. At the same time, it can also be comprehensively assessed by monitoring indicators such as CPU utilization, memory bandwidth, and I / O wait time, and combining the volatility and peak values ​​of these indicators to quantify the degree of resource contention within the system, providing key input for early warning and scheduling.The caching benefit coefficient is an indicator that evaluates the efficiency of system cache utilization and its contribution to overall performance. A high benefit coefficient usually indicates that the cache is configured and used properly, effectively improving system throughput and response speed. This coefficient can be calculated using parameters such as cache hit rate, current system throughput, and cache memory size. Furthermore, it can be comprehensively evaluated by analyzing cache read / write latency, cache miss counts, cache line replacement strategies, and other system performance indicators to assess cache benefit and evaluate the utilization efficiency of cache resources, guiding the scheduling module to consider cache optimization during resource allocation. The weighted fan-out coefficient measures the impact of data transmission volume and call frequency when a service node initiates calls to other service nodes on the overall system load. It reflects the complexity of inter-service communication and the overhead of data transmission. This coefficient can be calculated by summing the ratio of semantically compressed data transmission volume to a preset baseline data volume in each cross-service call. It can also be weighted by considering factors such as the importance of each call, call frequency, and real-time data transmission requirements to quantify the complexity and overhead of inter-service data transmission, and is an important component in calculating equivalent processing time. Equivalent processing time measures the total time a request takes from start to finish, taking into account the processing time of all service nodes, data transmission latency, and network overhead. It is a comprehensive performance metric that reflects the end-to-end latency of the entire request chain. This time can be calculated by summing the processing times of each service node and adding the data transmission latency calculated based on the semantically compressed data transmission volume, the available bandwidth of the root node, the network protocol overhead factor, and the weighted fan-out coefficient. Alternatively, end-to-end latency can be directly measured through a distributed link tracing system and combined with models for prediction and calibration, providing a unified metric to evaluate the overall response performance of the system. It is one of the key inputs for scheduling decisions.

[0116] The present invention quantifies the interdependencies between system variables by constructing a coupling degree matrix, wherein the elements in the matrix are, for example, The Pearson correlation coefficient represents the relationship between equivalent processing time and the resource competition index. The Pearson correlation coefficient represents the relationship between equivalent processing time and cache benefit factor. The Pearson correlation coefficient, representing the resource contention index and the cache benefit coefficient, is calculated based on sampled data from the most recent time window, and the coupling matrix is ​​a symmetric matrix. This quantification method allows the system to clearly identify the synergistic effects between different performance indicators; for example, how an increase in processing time might affect resource contention, or how an improvement in cache efficiency might conversely affect processing time. Based on this, the scheduling module further uses the formula for coupling entropy... The overall coupling degree of the computational system is calculated. Coupling entropy provides a single numerical value to characterize the complexity and uncertainty within the system. Higher coupling entropy means that the interactions between different parts of the system are more significant, requiring more careful and precise scheduling decisions. To transform this complex system state information into specific scheduling instructions, the scheduling module introduces a nonlinear scheduling decision factor. This factor comprehensively considers coupling entropy. Resource competition index and equivalent processing time Several key indicators, and through non-linear exponents Weighting is applied. This nonlinear combination allows decision factors to more sensitively capture subtle changes in the system state, thereby generating more adaptive and robust scheduling strategies. For example, when the resource contention index or equivalent processing time exceeds a preset threshold, the introduction of the nonlinear index can rapidly amplify these negative impacts, prompting the scheduling module to take more aggressive intervention measures. To accurately calculate the above factors, this scheme also defines in detail the acquisition methods of several core indicators. Resource Contention Index By comparing the number of instructions per clock cycle of the predictive model service process Number of instructions per clock cycle for the pricing engine service process The differences, combined with the number of access conflicts in the shared LLC cache per unit time. Total number of accesses to the shared LLC cache per unit time The ratio is used to quantify, where This represents the system's theoretical maximum number of instructions per clock cycle, thus accurately reflecting the degree of contention for computing resources by core services. Cache benefit coefficient. Based on cache hit rate Current system throughput and the memory usage of the cache To evaluate the utilization efficiency of cache resources, among which This represents the equivalent processing time, guiding the scheduling module to consider cache optimization during resource allocation. Furthermore, this scheme also introduces a weighted fan-out coefficient. and equivalent processing time The calculation of the weighted fan-out coefficient. After semantic compression, the first Data transfer volume per call Compared with the preset baseline data volume The sum of the ratios is obtained, where Indicates the number of cross-service calls. Equivalent processing time. This takes into account the total number of service nodes. , No. Processing time of each service node Available bandwidth of the root node Network protocol overhead factor and weighted fan-out coefficient This provides an end-to-end performance evaluation metric. The precise calculation of these metrics enables the scheduling module to comprehensively assess system performance bottlenecks and provide accurate input for nonlinear scheduling decision factors. Through this mechanism, the scheduling module no longer relies solely on the resource conflict probability predicted by the early warning module, but can deeply understand the interactions of internal system variables, quantify their complexity, and generate refined nonlinear scheduling decisions based on these quantification results. This allows the system to transform from passive conflict early warning to proactive, intelligent resource scheduling, thereby achieving better performance and stability in complex and ever-changing environments. Network protocol overhead factor and nonlinear exponent All of these can be obtained through expert experience assignment, historical data fitting, or experimental calibration.

[0117] The following is a concrete example to illustrate this. The scheduling module can periodically (e.g., every 5 seconds) execute a scheduling decision process. First, when constructing the coupling matrix, the system can continuously collect sample data on equivalent processing time, resource contention index, and cache benefit coefficient within the most recent 100 time windows. For example, equivalent processing time can be calculated from the service node processing time and inter-service data transfer volume obtained by the data acquisition module; the resource contention index can be obtained by monitoring the CPU performance counter and LLC cache controller logs; and the cache benefit coefficient can be obtained from the performance monitoring interface of the operating system or virtualization platform, such as cache hit rate and system throughput. Using this sample data, the scheduling module can calculate the Pearson correlation coefficients between equivalent processing time and resource contention index, equivalent processing time and cache benefit coefficient, and resource contention index and cache benefit coefficient, and fill them into the coupling matrix C. Next, the scheduling module will use this real-time updated coupling matrix to calculate the current coupling entropy according to a preset formula. For example, if the elements in the matrix It is 0.7. It is 0.5. If the value is 0.6, it can be substituted into the coupling entropy formula for calculation. Subsequently, the scheduling module will combine the currently calculated coupling entropy. Real-time monitored resource competition index (For example, currently 0.8) and equivalent processing time (For example, currently 150ms), and the preset maximum allowable coupling entropy. Maximum Permissible Resource Competition Index Equivalent processing time threshold and nonlinear exponent (For example, Set to 2) to calculate the nonlinear scheduling decision factor. For example, the resource competition index Specifically, this can be achieved by reading the CPU's performance monitoring unit registers to obtain the number of instruction retirements for the prediction model service process and the pricing engine service process within a specific time period, and then calculating... and Simultaneously, the number of access conflicts to the shared LLC cache is counted using APIs provided by the operating system or hardware performance counters. Total number of visits Then substitute into the formula to calculate. Caching benefit coefficient Specifically, this can be achieved by obtaining the current cache hit rate through system performance monitoring tools. (For example, 95%) system throughput (For example, 10,000 requests per second) and cache memory usage. (For example, 2GB), then substitute into the formula to calculate. Equivalent processing time Specifically, the data acquisition module obtains the processing time of each service node (e.g., M=3 nodes) from the distributed tracing system. (e.g., 10ms, 20ms, 15ms), and the amount of cross-service call data transferred after semantic compression. (For example, N=2 calls, with values ​​of 5KB and 8KB respectively). Simultaneously, obtain the available bandwidth of the root node. (e.g., 1Gbps) and network protocol overhead factor (For example, 1.2). Weighted fan-out coefficient Then by Compared with the preset baseline data volume The summation of ratios (e.g., 1KB) is obtained. Substituting these values ​​into the equivalent processing time formula yields the result. Finally, the scheduling module will make decisions based on the calculated nonlinear scheduling factors. Combined with a pre-defined scheduling strategy library, it dynamically selects and executes the most suitable resource scheduling strategy, such as adjusting the number of service instances, migrating service instances, adjusting resource quotas, or optimizing network paths, in order to cope with the current complex system state.

[0118] Through the above technical solutions, this system overcomes the limitations of traditional scheduling methods that rely solely on single indicators or simple rules. By constructing a coupling degree matrix and calculating coupling entropy, the system can gain a deep understanding of the complex interactions between key performance indicators such as equivalent processing time, resource contention index, and cache benefit coefficient, thereby avoiding new performance bottlenecks caused by one-sided optimization. The introduction of nonlinear scheduling decision factors enables the scheduling strategy to respond more accurately and flexibly to subtle changes in the system state, especially when resource contention is intense or processing time is close to the threshold, rapidly amplifying these negative impacts and prompting the system to take more proactive intervention measures. Furthermore, the precise quantification of the resource contention index, cache benefit coefficient, and equivalent processing time provides a solid data foundation for scheduling decisions, ensuring the scientific nature and effectiveness of the decisions. This allows the scheduling module to shift from passive fault response to proactive performance optimization, significantly improving the resource utilization efficiency, system stability, and user experience of the intelligent electronic component sales system under a complex microservice architecture. The system can more effectively balance performance, cost, and reliability, ensuring that underlying resources are optimally configured and utilized during demand forecasting and dynamic pricing.

[0119] Preferably, the rule by which the feedback optimization module updates the resource contention early warning threshold is:

[0120]

[0121] in, This indicates the updated resource contention warning threshold. This represents the rate of change of the equivalent processing time before and after scheduling. Indicates the threshold adjustment step size. This indicates the resource contention warning threshold before the update;

[0122] The rate of change of the equivalent processing time before and after scheduling is expressed as:

[0123]

[0124] in, This represents the rate of change of the equivalent processing time before and after scheduling. This represents the equivalent processing time before scheduling. This indicates the equivalent processing time after scheduling.

[0125] The rule for updating the resource contention warning threshold defines how to adjust the threshold based on the rate of change of equivalent processing time before and after scheduling. Its core lies in introducing a response mechanism to changes in system performance, enabling the warning threshold to adjust adaptively. When the rate of change of equivalent processing time deviates significantly from zero (whether significantly decreasing or significantly increasing), the system will correspondingly lower the warning threshold to detect potential resource contention problems earlier. This adjustment can be a multiplicative adjustment based on a preset adjustment step size; for example, a linear decay factor or an exponential decay factor can be used to calculate the new threshold. The rate of change of equivalent processing time before and after scheduling is a key indicator for measuring the improvement or deterioration of system performance before and after the execution of the scheduling strategy. It is calculated by comparing the equivalent processing time after scheduling with the equivalent processing time before scheduling. This indicator can intuitively reflect the effectiveness of the current scheduling strategy. For example, if the equivalent processing time is significantly shortened after scheduling, it indicates that the scheduling strategy is effective and system performance is improved; conversely, if the equivalent processing time increases, it may mean that the scheduling strategy has failed to achieve the expected results or even leads to performance degradation. The rate of change can be calculated using a relative change or the ratio of an absolute change to a baseline value. The threshold adjustment step size is a preset parameter used to control the magnitude of each adjustment to the resource contention warning threshold. This step size determines the system's response speed and stability to performance changes. A larger step size allows the threshold to adapt to system changes more quickly, but may also lead to excessive threshold fluctuations and unnecessary frequent adjustments; a smaller step size provides a more stable threshold, but its adaptability may be poor. This step size can be a fixed small value, such as 0.01 or 0.05, or it can be an adaptive value dynamically calculated based on historical system performance data.

[0126] The feedback optimization module of this invention collects performance data after the scheduling strategy is executed and dynamically updates the warning threshold and scheduling parameters based on this data. Specifically, after the scheduling strategy is executed, the feedback optimization module calculates the rate of change of the equivalent processing time before and after scheduling. Equivalent processing time is a key indicator for measuring system performance, and its rate of change directly reflects the actual effect of the scheduling strategy. When a significant deviation of the rate of change of the equivalent processing time from zero is detected (e.g., ...), < -0.1 or A threshold > 0.1 indicates a significant change in system performance, at which point the original resource contention warning threshold may no longer be applicable. To enable the warning system to better adapt to the current system state, the feedback optimization module adjusts the step size based on a preset threshold, thereby adjusting the resource contention warning threshold and generating a new threshold. This adjustment mechanism allows the warning threshold to adaptively decrease based on actual system performance feedback, thus improving the sensitivity of the warning and ensuring timely detection of potential resource contention issues when system performance fluctuates. In this way, the warning module no longer relies on static thresholds but can dynamically adjust based on the actual system performance, significantly improving the accuracy and real-time performance of the warnings, avoiding false alarms or missed alarms due to inappropriate thresholds, and ultimately optimizing the resource management efficiency and responsiveness of the entire intelligent sales system.

[0127] The following is a concrete example to illustrate this. The feedback optimization module can trigger its update logic periodically or after each execution of a scheduling policy. For example, after a resource scheduling policy is completed, the system will immediately collect the new equivalent processing time. and the equivalent processing time before scheduling. Compare and calculate the rate of change Assuming a preset threshold adjustment step size. It is 0.05. If the calculated value is... -0.15 (indicating a significant reduction in equivalent processing time and a performance improvement), or A value of 0.2 (indicating a significant increase in equivalent processing time and a performance decrease) indicates a significant change in system performance, necessitating adjustment of the warning threshold. In this case, the feedback optimization module will adjust the current resource contention warning threshold. Multiply by (1 - 0.05) to obtain the new warning threshold. For example, if If it is 0.8, then It will be updated to 0.8 (1-0.05) = 0.76. If Between -0.1 and 0.1 (e.g.) If the value is 0.02, then the system performance change is considered insignificant, and the warning threshold is set. This remains unchanged. This mechanism can be deployed as a standalone microservice or integrated into the post-processing logic of the scheduling module, receiving scheduling results and performance data via a message queue and asynchronously performing threshold update operations.

[0128] Through the above technical solution, the feedback optimization module of this invention can dynamically adjust the resource contention early warning threshold according to the actual system performance changes after the scheduling strategy is executed. This adaptive threshold adjustment mechanism enables the early warning module to more accurately reflect the current operating status of the system, avoiding the problem of inaccurate early warnings caused by using a fixed threshold. When system performance changes significantly due to scheduling, whether it improves or deteriorates, the early warning threshold will decrease accordingly, thereby improving the sensitivity of the early warning and enabling the system to detect potential resource contention risks more promptly. This not only reduces the occurrence of false alarms and missed alarms, but also ensures that the early warning system can continuously and effectively support the stable operation and efficient resource management of the intelligent sales system in a dynamic environment, thereby improving the robustness and responsiveness of the entire system.

[0129] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0130] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent sales system for electronic components based on demand forecasting and dynamic pricing, characterized in that, include: The data acquisition module collects microservice call chain data through a distributed tracing system. The call chain data includes service node processing time and cross-service call data transmission volume. The semantic compression module receives a high-dimensional tensor output by an external prediction model and performs feature compression on the high-dimensional tensor through an autoencoder neural network to generate a low-dimensional feature vector that the pricing engine can directly consume, thereby eliminating runtime format conversion overhead. The early warning module, based on the historical resource competition index sequence, uses a time-series prediction model to predict the probability of resource conflicts within future time windows; The scheduling module constructs the coupling degree matrix of system variables, obtains the coupling entropy, and executes resource scheduling strategies based on nonlinear scheduling decision factors. The feedback optimization module collects performance data after the scheduling strategy is executed and dynamically updates the warning threshold and scheduling parameters. The autoencoder includes an encoder and a decoder. The encoder consists of multiple fully connected layers and nonlinear activation functions, used to map high-dimensional inputs to low-dimensional latent spaces. The decoder consists of multiple fully connected layers and nonlinear activation functions, used to reconstruct low-dimensional latent variables into the original high-dimensional space.

2. The intelligent sales system for electronic components based on demand forecasting and dynamic pricing according to claim 1, characterized in that, The loss function used by the semantic compression module is: ； in, This indicates the loss function used by the semantic compression module. This represents a high-dimensional tensor received from an external prediction model. The encoder function, consisting of multiple fully connected layers and nonlinear activation functions, maps the input to a low-dimensional latent space. The decoder function, consisting of multiple fully connected layers and non-linear activation functions, reconstructs the latent variables to their original dimensions. Represents the square of the L2 norm. Describing the L1 norm, This represents the sparse regularization coefficient.

3. The intelligent sales system for electronic components based on demand forecasting and dynamic pricing according to claim 1, characterized in that, The predictive early warning module uses a Long Short-Term Memory (LSTM) network as the time-series prediction model, and inputs historical data. The resource competition index sequence for each time window outputs the future... The predicted value for each time window; the formula for calculating the probability of resource conflict is: ； in, This indicates the probability of resource conflicts occurring within the future window. This represents the future output of the LSTM model. The predicted value of the resource competition index for each time window. This represents the preset resource competition warning threshold. This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

4. The intelligent sales system for electronic components based on demand forecasting and dynamic pricing according to claim 2, characterized in that, The coupling matrix constructed by the scheduling module is as follows: ； in, Represents the coupling degree matrix. The Pearson correlation coefficient represents the relationship between equivalent processing time and the resource competition index. The Pearson correlation coefficient represents the relationship between equivalent processing time and cache benefit factor. The Pearson correlation coefficient represents the relationship between the resource contention index and the cache benefit coefficient. Each Pearson correlation coefficient is calculated based on the sampling data of the most recent time window, and the coupling matrix is ​​a symmetric matrix.

5. The intelligent sales system for electronic components based on demand forecasting and dynamic pricing according to claim 4, characterized in that, The formula for obtaining the coupling entropy in the scheduling module is: ； in, Represents coupling entropy, This represents an element in the coupling degree matrix.

6. The intelligent sales system for electronic components based on demand forecasting and dynamic pricing according to claim 5, characterized in that, The formula for obtaining the nonlinear scheduling decision factor in the scheduling module is as follows: ； in, Represents the nonlinear scheduling decision factor. Represents coupling entropy, Maximum allowable coupling entropy This represents the resource competition index. This represents the maximum permissible resource competition index. Indicates the equivalent processing time. Indicates the equivalent processing time threshold. This represents a nonlinear exponent.

7. The intelligent sales system for electronic components based on demand forecasting and dynamic pricing according to claim 6, characterized in that, The formula for obtaining the resource competition index is: ； in, This represents the resource competition index. This represents the number of instructions per clock cycle for the predictive model service process. This represents the number of instructions per clock cycle for the pricing engine service process. This represents the system's theoretical maximum number of instructions per clock cycle. This represents the number of access conflicts in the shared LLC cache per unit of time. This represents the total number of accesses to the shared LLC cache per unit of time.

8. The intelligent sales system for electronic components based on demand forecasting and dynamic pricing according to claim 6, characterized in that, The cache benefit coefficient is obtained as follows: based on the cache hit rate, the current system throughput, and the cache memory size, the cache benefit coefficient is obtained using the following formula: ； in, This represents the caching benefit coefficient. Indicates cache hit rate. This indicates the current system throughput. This indicates the memory size occupied by the cache. Indicates the equivalent processing time.

9. The intelligent sales system for electronic components based on demand forecasting and dynamic pricing according to claim 6, characterized in that, The equivalent processing time is obtained as follows: Based on the semantically compressed first The weighted fan-out coefficient is obtained by using the following formula, based on the amount of data transferred in each call and the preset baseline data amount: ； in, This represents the weighted fan-out coefficient. Indicates the number of cross-service calls. Indicates the semantically compressed first... Data transfer volume per call Indicates the preset baseline data volume; Based on the The equivalent processing time is obtained by considering the processing time of each service node, the available bandwidth of the root node, and the weighted fan-out coefficient, using the following formula: ； in, Indicates the equivalent processing time. Indicates the total number of service nodes. Indicates the number of cross-service calls. Indicates the first Processing time of each service node Available bandwidth for the root node Indicates the semantically compressed first... Data transfer volume per call Indicates the network protocol overhead factor. This represents the weighted fan-out coefficient.

10. The intelligent sales system for electronic components based on demand forecasting and dynamic pricing according to any one of claims 1-9, characterized in that, The rule for the feedback optimization module to update the resource contention early warning threshold is as follows: ； in, This indicates the updated resource contention warning threshold. This represents the rate of change of the equivalent processing time before and after scheduling. Indicates the threshold adjustment step size. This indicates the resource contention warning threshold before the update; The rate of change of the equivalent processing time before and after scheduling is expressed as: ； in, This represents the rate of change of the equivalent processing time before and after scheduling. This represents the equivalent processing time before scheduling. This indicates the equivalent processing time after scheduling.

Images