New energy industry supply chain monitoring and early warning method and system based on behavior pattern

Abstract

The present application relates to the technical field of supply chain data processing and monitoring and early warning, and discloses a new energy industry supply chain monitoring and early warning method and system based on behavior patterns, which comprises the following steps: acquiring a bottom physical sequence and a surface transaction sequence, extracting a first derivative to generate a phase feature flow, aligning a historical sequence to generate a standard benchmark phase matrix; inputting environmental and equipment depreciation parameters into a time lag compensation model to calculate a legal time lag increment, combining the benchmark matrix to generate a dynamic reference phase matrix, and solving a phase deviation degree coefficient; when the deviation degree coefficient reaches a control limit, injecting a perturbation label into a purchase message, comparing actual and expected response gradients to obtain a response consistency score; finally, weighting the deviation degree coefficient and the response consistency score to obtain a comprehensive early warning intensity index, and determining an abnormal behavior pattern. The present application can effectively strip the physical delay caused by the objective environment and equipment aging, actively verify the real execution capability of the production line, and improve the accuracy and reliability of supply chain anomaly monitoring.

Description

Technical Field

[0001] This invention relates to the field of supply chain data processing and monitoring and early warning technology, specifically to a method and system for monitoring and early warning of the new energy industry supply chain based on behavioral patterns. Background Technology

[0002] The stability of the new energy industry supply chain is directly related to the on-time delivery of overall production capacity. The new energy industry chain involves complex manufacturing processes, and the companies at each node not only have surface-level order transactions and information flow, but also underlying physical processes of energy consumption and material throughput.

[0003] Currently, most supply chain monitoring and early warning solutions rely on surface-level transaction data from business systems. Existing monitoring methods typically only compare order placement time with node feedback time, failing to delve into the underlying physical production status at the workshop level. When delays or logical errors occur in surface-level business data, the system struggles to distinguish between information network congestion and actual production line shutdowns. Furthermore, existing technologies often use fixed time thresholds to determine delivery defaults or abnormal behavior, completely ignoring the impact of environmental temperature fluctuations and wear and tear on core production line equipment during actual manufacturing operations. These objective factors can slow down production line processing cycles and create reasonable physical delays. Attributing all delays directly to anomalies without considering these factors easily leads to frequent false alarms from the early warning system.

[0004] In addition, existing monitoring methods are generally in a passive state of receiving data. When early signs of data deviation are detected, they cannot actively issue probing instructions to the target node to verify its actual production response capability. As a result, the final anomaly judgment results lack physical-level cross-verification, and the overall early warning reliability is difficult to meet the needs of refined management and control. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a new energy industry supply chain monitoring and early warning method and system based on behavioral patterns. This solves the problems of high false alarm rate and low reliability of existing supply chain monitoring technologies, which rely solely on surface transaction data and ignore objective physical delays caused by environmental fluctuations and equipment aging, as well as the lack of proactive verification methods for the underlying real production status.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a method for monitoring and early warning of the new energy industry supply chain based on behavioral patterns, the method comprising the following steps: S100: Obtain the underlying physical sequence and the surface transaction sequence, extract the first derivative to generate the current phase feature stream; align the historical phase feature stream with the historical surface transaction sequence, calculate the median time offset, and generate the standard reference phase matrix; S200 collects environmental state parameters and equipment depreciation state parameters, inputs them into a time delay compensation neural network model, calculates the legal time delay increment, adds it to the standard reference phase matrix to generate a dynamic reference phase matrix; calculates the current phase feature flow and the surface transaction sequence delay to construct the actual phase difference matrix; calculates the residual norm of the actual phase difference matrix and the dynamic reference phase matrix to solve for the phase deviation coefficient; When the phase deviation coefficient of S300 reaches the first-level control limit, the procurement message is injected with a perturbation tag and issued; the actual response gradient of the underlying physical sequence is collected, and the inner product of the actual response gradient and the expected response gradient is calculated to obtain the response consistency score. S400 calculates a comprehensive early warning intensity index by weighting the phase deviation coefficient and the response consistency score, determines abnormal behavior patterns, and outputs control instructions.

[0008] Preferably, in step S100, the underlying physical sequence includes a kinetic energy flow sequence characterizing the electrical load, a waste discharge sequence characterizing the sewage discharge situation, and a material throughput sequence characterizing the material inflow and outflow; the extraction of the first derivative to generate the current phase feature flow specifically includes: The underlying physical sequence is aligned to a uniform discrete time step using an interpolation algorithm; The first derivative of the aligned underlying physical sequence in the time dimension is calculated, and the absolute amplitude parameter is discarded by applying a sign function to generate the current phase feature stream composed of discrete state parameters.

[0009] Preferably, in step S100, the step of aligning the historical phase feature stream with the historical surface transaction sequence, calculating the median time offset, and generating a standard reference phase matrix specifically includes: Construct a two-dimensional distance matrix and calculate the local Euclidean distance between the historical phase feature flow sampling points and the historical surface transaction sequence sampling points, and solve for the regular path with the minimum cumulative distance; Extract multiple sets of time offsets from at least five complete historical delivery cycles; The statistical median of multiple time offset sets is calculated as the inherent physical time delay, and a standard reference phase matrix is ​​constructed by assembling the inherent physical time delays.

[0010] Preferably, in step S200, the environmental state parameters include the plant ambient temperature and the power grid frequency deviation, and the equipment depreciation state parameters include the cumulative total operating time of the core production line and the spindle calibration wear degree; the specific steps of collecting the environmental state parameters and equipment depreciation state parameters and inputting them into the time delay compensation neural network model to calculate the legal time delay increment include: The environmental state parameters and equipment depreciation state parameters are concatenated and input into a time-delay compensation neural network model built on a multilayer perceptron architecture. The nonlinear cross-coupling effect contained in the parameters is extracted by a fully connected hidden layer, and the time delay elongation caused by temperature and wear is calculated by the output layer as a legitimate time delay increment.

[0011] Preferably, in step S200, the calculation of the residual norm of the actual phase difference matrix and the dynamic reference phase matrix to solve for the phase deviation coefficient specifically includes: Calculate the difference matrix between the actual phase difference matrix and the dynamic reference phase matrix; Solve for the L2 norm contained in the difference matrix, divide the solved L2 norm by the sum of the L2 norm contained in the dynamic reference phase matrix and the preset smoothing positive real constant, and obtain the dimensionless phase deviation coefficient. When there is a dimensional misalignment between the incremental matrix and the standard reference phase matrix due to missing sampling, zero-filling or dimensionality reduction truncation based on singular value decomposition is performed in advance to force the alignment of the two dimensions.

[0012] Preferably, in step S300, the injection of the purchase message perturbation tag specifically includes: When the phase deviation coefficient continuously exceeds the first-level control limit set based on three times the standard deviation of the historical residual mean within the set observation time window, the procurement message to be issued will be intercepted. A perturbation tag is dynamically written into the extended protocol field included in the procurement message. The perturbation tag is used to instruct the target node to make minor adjustments to specific process parameters of the current batch of raw materials. The adjustment range is set between 0.1% and 0.5% of the rated operating parameters.

[0013] Preferably, in step S300, the acquisition of the actual response gradient of the underlying physical sequence specifically includes: Within the predetermined response time window after the message reaches the target gateway, a configuration update command is sent to the edge computing gateway to force an increase in the sampling frequency of the underlying physical sequence; Extract the underlying physical sequence within the high-frequency sampling period, and extract the actual response gradient vector in the time domain through first-order difference operation; The production line digital twin process library is invoked to substitute the perturbation amplitude into the equipment dynamics polynomial to generate the expected response gradient.

[0014] Preferably, in step S400, the weighted average of the phase deviation coefficient and the response consistency score to obtain the comprehensive early warning intensity index specifically includes: Extract the current production order business sensitivity index from the enterprise resource planning system. The business sensitivity index is obtained by linearly summing the order cash flow amount and the delivery urgency after performing maximum and minimum normalization. The steady-state deviation penalty term included in the phase deviation coefficient is calculated using a logarithmic function, and the dynamic trial penalty term included in the response consistency score is calculated using an inverse proportional function. A comprehensive early warning intensity index is generated by performing a linear combination weighted average of the steady-state deviation penalty term, the dynamic trial penalty term, and the business sensitivity index.

[0015] Preferably, in step S400, determining the abnormal behavior pattern specifically includes: The phase deviation coefficient, response consistency score, and comprehensive early warning intensity index are concatenated and integrated into a three-dimensional state feature vector. The mean-variance standardization operation is used to perform dimensional alignment on the three-dimensional state feature vectors. The aligned 3D state feature vectors are fed into a behavior pattern mapping neural network containing a multi-layer feedforward perceptron architecture, and the classification probability distribution is output through a normalized exponential function. The pattern corresponding to the maximum value in the classification probability distribution is selected as the final qualitative conclusion of the anomaly, and it is classified as a behavioral anomaly pattern.

[0016] A second aspect of this invention provides a new energy industry supply chain monitoring and early warning system based on behavioral patterns, the system comprising: The multidimensional feature perception and benchmark self-learning module includes an edge feature desensitization unit and a benchmark matrix generation unit. The edge feature desensitization unit is used to obtain the underlying physical and surface transaction sequences and extract the first derivative to generate the current phase feature stream. The benchmark matrix generation unit is used to align the historical sequences, calculate the median time offset, and generate a standard benchmark phase matrix. The adaptive modulation and deviation calculation module includes an environmental equipment coupling modulation unit and a phase deviation calculation unit. The environmental equipment coupling modulation unit is used to calculate the legal time delay increment based on environmental and equipment depreciation status parameters and generate a dynamic reference phase matrix. The phase deviation calculation unit is used to calculate the phase deviation coefficient by calculating the residual norm of the current actual phase difference matrix and the dynamic reference phase matrix. The perturbation probe targeting verification module includes a message probe injection unit and a physical response verification unit. The message probe injection unit is used to inject perturbation tags into the procurement message when the phase deviation coefficient reaches the first level control limit. The physical response verification unit is used to collect the underlying actual response gradient and calculate its inner product with the expected response gradient to obtain the response consistency score. The integrated decision-making and control execution module includes an early warning intensity calculation unit and an abnormal mode execution unit. The early warning intensity calculation unit is used to calculate the comprehensive early warning intensity index by weighting the deviation coefficient and the response consistency score. The abnormal mode execution unit is used to determine the abnormal behavior mode and output control instructions.

[0017] This invention provides a method and system for monitoring and early warning of the new energy industry supply chain based on behavioral patterns. It has the following beneficial effects: 1. This invention generates a phase feature stream by extracting the first derivatives of the physical and transaction sequences, and constructs a standard reference phase matrix by calculating the median time offset. This feature removes the differences in absolute amplitude across modal data, retaining only the direction of change, objectively reflecting the temporal synchronization status of material flow and book transactions, and providing an accurate basis for deviation calculation.

[0018] 2. This invention calculates the legitimate time delay increment by inputting environmental and equipment depreciation parameters into a time delay compensation neural network, thereby generating a dynamic reference phase matrix. This feature incorporates objective physical hysteresis caused by temperature fluctuations or machine wear into the deviation calculation, effectively eliminating delays caused by normal equipment aging and significantly reducing the false alarm rate.

[0019] 3. This invention injects a perturbation tag into the procurement message when the phase deviation coefficient reaches the control limit, and calculates the inner product of the actual and expected response gradients to obtain a response consistency score. This mechanism directly verifies the physical execution capability of the production line by issuing minor process adjustment instructions, thereby accurately distinguishing between data delays and substantial production stoppages, and improving the reliability of anomaly detection. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall process of the new energy industry supply chain monitoring and early warning method based on behavioral patterns according to the present invention. Figure 2 This is a block diagram of the module structure of the new energy industry supply chain monitoring and early warning system based on behavioral patterns according to the present invention. Figure 3 This is a schematic diagram illustrating the principle of multi-source feature edge desensitization and reference phase matrix generation in this invention. Figure 4 This is a flowchart of the environmental device coupling adaptive modulation and deviation calculation process of the present invention; Figure 5 This is a schematic diagram of the process for perturbation probe-based targeted verification of the present invention; Figure 6 This is a flowchart of the integrated early warning intensity weighted solution and behavior anomaly pattern mapping of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 , Figure 1This is a general flowchart of a new energy industry supply chain monitoring and early warning method based on behavioral patterns according to an embodiment of the present invention. The present invention provides a new energy industry supply chain monitoring and early warning method based on behavioral patterns, which is applied to a new energy industry supply chain monitoring and early warning system based on behavioral patterns. The method may include the following steps: S100: Obtain the underlying physical sequence and surface transaction sequence of the supply chain node in the current cycle. At the edge, perform gradient masking to desensitize the underlying physical sequence, extract the first derivative while preserving the direction of change, and generate a current phase feature stream with timestamps. Retrieve the historical phase feature stream and historical surface transaction sequence of the supply chain node within its historical steady-state cycle, and use a dynamic time warping algorithm to align the historical phase feature stream and historical surface transaction sequence with optimal paths. Generate a standard reference phase matrix by calculating the median time offset of multiple aligned paths.

[0023] In S200, environmental state parameters and equipment depreciation state parameters of supply chain nodes are collected to construct an environmental-equipment coupling matrix. This coupling matrix is ​​then input into a pre-trained time-delay compensation neural network model to capture nonlinear physical cross-coupling effects and calculate the legitimate time-delay increment caused by objective operating condition deterioration. This legitimate time-delay increment is added to the standard reference phase matrix generated in S100 to generate a dynamic reference phase matrix for comparison in the current cycle. Based on the cross-correlation function, the abrupt edge delay between the current phase feature flow generated in S100 and the surface transaction sequence is calculated to construct the actual phase difference matrix. The residual norm of the actual phase difference matrix and the dynamic reference phase matrix is ​​extracted to calculate the phase deviation coefficient.

[0024] S300 determines whether the phase deviation coefficient calculated in S200 reaches the preset first-level control limit. When the phase deviation coefficient reaches the first-level control limit, an active probing mechanism is triggered. A perturbation tag is attached and injected into the extended field of the regular electronic data interchange procurement message, and the message carrying the perturbation tag is sent to the corresponding supply chain node. Within a preset time window after the message is sent, the actual response gradient of the underlying physical sequence of the node in response to the perturbation tag is collected frequently. The inner product of the actual response gradient and the expected physical response gradient under the preset process model is calculated to obtain the response consistency score.

[0025] In step S400, the phase deviation coefficient calculated in step S200 and the response consistency score obtained in step S300 are compared. A weighted average of the phase deviation coefficient and the response consistency score is then calculated to obtain the comprehensive early warning intensity index. Based on the joint distribution results of the comprehensive early warning intensity index and the phase deviation coefficient, the abnormal behavior patterns of the current supply chain nodes are determined. These abnormal behavior patterns are categorized into two types: a capacity stagnation pattern (representing false capacity) and a yield failure pattern (representing physical production line malfunctions). According to the specific abnormal behavior pattern identified, corresponding supply chain capacity transfer or control instructions are output through the enterprise service bus.

[0026] See attached document Figure 2 , Figure 2 This is a module structure diagram of a new energy industry supply chain monitoring and early warning system based on behavioral patterns according to an embodiment of the present invention. The present invention provides a new energy industry supply chain monitoring and early warning system based on behavioral patterns, which includes a multi-dimensional feature perception and benchmark self-learning module, an adaptive modulation and deviation calculation module, a perturbation probe targeted verification module, and a comprehensive decision-making and control execution module.

[0027] The multidimensional feature perception and benchmark self-learning module is used to perform data acquisition, feature desensitization, and benchmark model bootstrapping calibration, specifically including: Edge Feature Desensitization Unit: Deployed at the on-site edge computing gateway, it acquires the underlying physical sequence containing kinetic energy flow, waste discharge, and logistics throughput, as well as the surface transaction sequence. After aligning the sequences to a uniform time step using an interpolation algorithm, it extracts the first derivative and applies a sign function to discard absolute amplitudes, generating a current phase feature stream that retains the trend of change. The mathematical process for calculating the current phase feature stream is shown in the formula: ; In the formula, This represents the phase characteristic flow vector at the current moment; This represents a three-dimensional underlying physical sequence feature vector composed of kinetic energy flow, waste discharge, and material throughput. This represents the feature vector of the underlying physical sequence at the previous sampling time. This indicates the sampling time interval, which is typically between 1 minute and 15 minutes, depending on the minimum physical cycle time of the production process. This represents a symbolic function. Through this differential processing, the system shifts its focus from the absolute quantity of output to the sequence of production actions, thus preserving behavioral characteristics while protecting trade secrets.

[0028] The benchmark matrix generation unit extracts feature flows and transaction sequences within historical steady-state cycles and uses a dynamic time warping algorithm to match the optimal alignment path. It calculates the median of the corresponding time offset as the inherent physical time delay, thereby constructing a standard benchmark phase matrix.

[0029] The adaptive modulation and deviation calculation module is used to correct the early warning benchmark based on fluctuations in the objective environment and to quantify the degree of abnormal deviation, specifically including: Environmental equipment coupling modulation unit: Collects environmental parameters such as ambient temperature and power grid frequency deviation, as well as equipment depreciation parameters such as total production line operating time, to construct a coupling matrix. This matrix is ​​then input into a time-delay compensation neural network to calculate the legal time-delay increment, and added to the standard reference phase matrix to generate a dynamic reference phase matrix. The specific operational relationships for generating the dynamic reference phase matrix are shown in the formula: ; In the formula, Represents the dynamic reference phase matrix; Represents the standard reference phase matrix; Indicates environmental condition parameters, such as temperature deviation in degrees Celsius; Indicates equipment depreciation status parameters, such as cumulative operating time in hours; This represents a polynomial regression function, whose coefficients are obtained by fitting historical normal operation data using the least squares method.

[0030] Phase deviation calculation unit: Calculates the true time delay between the current feature stream and the abrupt edge of the transaction sequence using the cross-correlation function, assembling it into an actual phase difference matrix. The phase deviation coefficient is obtained by solving the L2 norm of the actual and dynamic reference phase difference matrices. If the matrix dimensions are inconsistent, zero-padding or truncation is automatically performed.

[0031] The perturbation probe targeting verification module is used to proactively issue commands to verify the actual response capability when an initial risk is detected. Specifically, it includes: Message probe injection unit: When the phase deviation coefficient triggers the first level control limit, it intercepts the procurement message and writes a perturbation tag in the extended field, requiring the node to perform a slight adjustment of 0.1% to 0.5% to the process parameters to induce physical changes.

[0032] Physical Response Verification Unit: Within the response time window, the sampling frequency is increased to obtain the actual response gradient, and the expected physical response gradient under perturbation conditions is retrieved. The inner product of the two is calculated and normalized, outputting a response consistency score. The calculation process is shown in the formula: ; In the formula, This represents the response consistency score, with a value ranging from 0 to 1; This represents the actual response gradient vector of the underlying physical sequence; This represents the expected physical response gradient vector; This represents a preset minimum positive number to prevent calculation anomalies where the denominator is zero when the underlying physical system is completely unresponsive. If the calculation result is close to 1, it indicates that the physical production line has actually executed the perturbation command; if it is close to 0, it is determined to be a false response.

[0033] The integrated decision-making and control execution module is used to integrate previous indicators to define violations and coordinate control measures, specifically including: Early warning intensity calculation unit: Based on the risk level of nodes in the supply chain, the deviation coefficient and the response consistency score are linearly weighted to output a comprehensive early warning intensity index.

[0034] The abnormal mode execution unit integrates deviation, consistency score, and warning intensity into a three-dimensional feature vector, inputs it into a behavioral pattern mapping neural network, and determines the physical status of business operations, such as compliant production, recorded playback attacks, equipment idling forgery, or illegal outsourcing. It ultimately generates control instructions, including order freezing or backup supplier activation, and sends them to the scheduling system.

[0035] See attached document Figure 3 , Figure 3 This is a schematic diagram illustrating the principle of multi-source feature edge desensitization and reference phase matrix generation according to an embodiment of the present invention. In this embodiment, the multi-dimensional feature perception and reference self-learning module collaboratively executes a reference bootstrapping mechanism based on masking and sequence regularization on the edge computing side and the cloud. This mechanism transforms multi-source heterogeneous industrial underlying physical sampling data into phase features with absolute values ​​removed, and uses historical steady-state data to establish an objective time-delay reference model, providing a benchmark anchor point for subsequent anomaly monitoring.

[0036] S110, the edge feature desensitization unit acquires the underlying physical sequence and surface transaction sequence of the target supply chain node. The underlying physical sequence includes a kinetic energy flow sequence representing electrical load, a waste discharge sequence representing pollution levels, and a material throughput sequence representing material inflow and outflow. To address the clock asynchrony of multi-source data, the edge feature desensitization unit uses an interpolation algorithm to align the underlying physical sequence to a unified discrete time step.

[0037] S120, based on the timestamp-aligned data described above, to prevent the leakage of the target node's true absolute production capacity data during cloud aggregation, the edge feature desensitization unit performs gradient mask desensitization processing on the underlying physical sequence. The core logic of this process is to strip away the absolute amplitude reflecting the production volume, extracting only the differential directions reflecting the start / stop and strength / weakness evolution of production actions. By calculating the first derivative of the underlying physical sequence in the time dimension and applying a sign function, the edge feature desensitization unit generates a phase feature stream containing only discrete state vectors. Considering that direct differential calculation might result in a zero denominator error when sensor data stagnates or transmission is congested, the specific gradient mask desensitization calculation process is shown in the formula: ; In the formula, This represents the phase feature stream vector output at the current sampling time; This represents the three-dimensional underlying physical sequence feature vector at the current moment, consisting of kinetic energy flow, waste discharge, and material throughput. This represents the feature vector of the underlying physical sequence at the previous sampling time. This represents the uniform sampling time interval after data alignment; This represents a preset, extremely small positive real number, whose value is usually set to 1. It is specifically designed to prevent division operation overflow crashes caused by a short-term sensor disconnection resulting in a zero time interval reading; The sign function maps the calculated positive gradient to positive 1, the negative gradient to negative 1, and the zero gradient to zero. The purpose of this masking technique is that the edge feature desensitization unit forces the mapping of continuous absolute values ​​to a low-dimensional discrete state space composed of -1, 0, and 1. This physically cuts off the path of inferring the company's actual output from intercepted data, while still accurately preserving the temporal characteristics of machine startup, material flow, and waste discharge.

[0038] S130, after obtaining the aforementioned desensitized phase feature stream, the system needs to construct a time delay benchmark for comparison for this supply chain node through the benchmark matrix generation unit. The benchmark matrix generation unit retrieves the historical phase feature stream and historical surface transaction sequence of this node within a time window of no historical defaults and stable production capacity. In this embodiment, the time window of stable production capacity is defined as a production cycle in which the monthly output fluctuation does not exceed 5% within three consecutive months. In order to eliminate the local time axis stretching and deformation caused by production scheduling in industrial production, the benchmark matrix generation unit introduces a dynamic time warping algorithm to elastically match the historical phase feature stream and the historical surface transaction sequence. The dynamic time warping algorithm achieves optimal alignment between the underlying physical behavior peak and the surface transaction confirmation peak by constructing a two-dimensional distance matrix and finding the warping path with the minimum cumulative distance. According to the dynamic programming principle, the recursive process of the elements of the cumulative distance matrix is ​​as shown in the formula: ; In the formula, This indicates the location of the grid point in the distance matrix reached by the regularized path. Minimum cumulative distance at; The first characteristic stream representing the historical phase flow Data from each sampling point The first of the historical surface transaction sequences Data from each sampling point Local Euclidean distance between them; This represents the function that takes the minimum value. To ensure the mathematical completeness of the above matrix recursive operations, the boundary conditions for initializing the reference matrix generating unit are as follows: and all and All values ​​are set to positive infinity. This recursive formula shows that the cumulative distance of the current coordinate node depends on its three adjacent historical state paths, thus allowing the sequence to be locally stretched or compressed on the time axis to overcome the defect that rigid point-to-point distance calculation cannot match misaligned peaks.

[0039] S140, combining the optimal alignment path obtained above, the benchmark matrix generation unit extracts multiple sets of time offsets from at least five complete historical delivery cycles; calculates the statistical median contained in the multiple sets of time offsets as the inherent physical time delay, and constructs a standard benchmark phase matrix by assembling the inherent physical time delays. Using the statistical median can effectively eliminate abnormal outliers caused by occasional queue jumping. This standard benchmark phase matrix is ​​permanently stored in the cloud-based early warning center for real-time high-frequency comparison and retrieval.

[0040] See attached document Figure 4 , Figure 4 This is a flowchart of the environmental equipment coupling adaptive modulation and deviation calculation process according to an embodiment of the present invention. In this embodiment, the adaptive modulation and deviation calculation module includes an environmental equipment coupling modulation unit and a phase deviation calculation unit. This structure is based on the prior reference bootstrapping mechanism, and dynamically corrects the static reference by introducing external objective operating condition variables, avoiding misjudging natural performance drift caused by changes in the physical environment as a violation.

[0041] S210, the environmental equipment coupling modulation unit acquires the environmental status parameters and equipment depreciation status parameters of the plant area in real time. To solve the problem of clock asynchrony and sampling frequency differences of multi-source data, the acquired parameters are resampled and aligned to a unified discrete time step using the sliding window averaging method. After outlier filtering and standardization preprocessing, the parameters are spliced ​​together to construct the environmental equipment coupling matrix.

[0042] S220, to accurately quantify the impact of working conditions on production rhythm, incorporates a time-delay compensation neural network model based on a multilayer perceptron architecture. This model takes the environmental equipment coupling matrix as input and outputs a vector of legitimate time-delay increments caused by deterioration of objective working conditions. During training, a loss function is constructed using mean squared error combined with a regularization penalty term to prevent overfitting, as shown in the following formula: ; In the formula, This represents the total training loss value; This represents the total number of samples in the training batch; This indicates the label representing the actual time delay difference obtained from parsing historical real data; This represents the predicted value of the legal time delay increment output by the forward propagation of the time delay compensation neural network model; This represents the regularization coefficient, which is usually set between 0.001 and 0.01. Its specific value is determined based on the network's generalization error performance during cross-validation. This represents the set of all weight parameter matrices in the network model; This represents a single connection weight variable.

[0043] S230, after completing the inference of the valid time-delay increment vector, the environmental device coupling modulation unit converts the valid time-delay increment vector into an increment matrix with the same dimension as the reference matrix, and performs element-wise matrix addition with the standard reference phase matrix to generate a dynamic reference phase matrix for comparison in the current period. The specific calculation process for generating the dynamic reference phase matrix is ​​shown in the formula: ; In the formula, Represents the dynamic reference phase matrix at the current moment; This represents the standard baseline phase matrix constructed during the previous stage of self-learning. This represents the valid time-delay increment matrix output by the time-delay compensated neural network model at the current moment, after dimensional expansion. To ensure the mathematical completeness of matrix addition operations, if there is a dimensional misalignment between the increment matrix and the standard reference phase matrix due to sampling omissions, the environmental equipment coupling modulation unit will pre-perform zero-padding or dimensionality reduction truncation operations based on singular value decomposition to forcibly align the dimensions of the two, preventing underlying operational anomalies such as array out-of-bounds errors. The purpose of the above dynamic modulation technique is to provide an objective judgment threshold benchmark for the anomaly monitoring system that can dynamically and adaptively adjust with seasonal changes and equipment aging.

[0044] S240, based on the aforementioned dynamically corrected benchmark, the phase deviation calculation unit is responsible for quantifying the true degree of anomaly in the current production cycle. The phase deviation calculation unit uses a time-domain cross-correlation function to perform a sliding convolution on the real-time acquired phase feature stream and the surface transaction sequence, searching for the time lag that maximizes the cross-correlation response amplitude, and combining them into the actual phase difference matrix. Subsequently, the phase deviation calculation unit calculates the relative residual L2 norm between the actual phase difference matrix and the dynamic reference phase matrix, outputting the phase deviation coefficient. To overcome the one-sidedness of single-dimensional indicator fluctuations, this calculation uses a global matrix norm form instead of comparing a single maximum value. The specific calculation process of the phase deviation coefficient is shown in the formula: ; In the formula, This represents the phase deviation coefficient, which is a dimensionless positive real number. This represents the actual phase difference matrix calculated within the current production cycle; This represents the dynamic reference phase matrix after operating condition compensation; This represents the L2 norm operation for solving matrices; This represents a preset smoothing positive real constant. The technical purpose of introducing this semi-slippery demand coefficient is to effectively avoid division overflow anomalies caused by the denominator being zero when the production line is in an extreme shutdown state causing the dynamic reference phase matrix to approach the zero vector. Its typical value is set between 0.01 and 0.1. The larger the phase deviation coefficient, the stronger the suspicion of purely human violation after the current supply chain node has deviated from the influence of the objective physical environment. In actual deployment, the system does not directly rely on a single extreme value for judgment, but presets a deviation tolerance control limit. An anomaly is only considered to exist when the deviation from the control limit is consistently exceeded within a consecutive test time window. This deviation tolerance control limit is typically based on 3 sigma (i.e., 3) calculated from historical normal operating conditions. The upper limit of the confidence interval is calibrated, thereby realizing a multi-dimensional weighted judgment logic that resists sudden noise interference at the algorithm level.

[0045] See attached document Figure 5 , Figure 5 This is a schematic diagram of a perturbation probe-based targeted verification process according to an embodiment of the present invention. In this embodiment, the perturbation probe-based targeted verification module is built upon the preliminary deviation calculation and is used to proactively issue test commands and verify the actual response capability of the underlying production line when the system determines that there is an initial risk of violation. This module specifically includes a message probe injection unit and a physical response verification unit, which, through a hardware-software collaborative probing mechanism, identifies whether there are fraudulent outsourcing or data replay attacks at supply chain nodes.

[0046] S310, the message probe injection unit monitors the phase deviation coefficient obtained from the aforementioned calculation in real time. When this coefficient continuously exceeds the first-level control limit within the set observation time window, the message probe injection unit intercepts the regular electronic data interchange procurement message that is about to be sent to the corresponding supply chain node. As a preferred method, the first-level control limit is set based on three times the standard deviation of the historical residual mean. After the interception is triggered, the message probe injection unit dynamically writes a perturbation tag into the extended protocol field of the procurement message through the protocol extension layer. Specifically, the lower-level feature is that the perturbation tag instructs the supply chain node to perform a non-standardized micro-adjustment of a specific process parameter of the current batch of raw materials. In this embodiment, the aforementioned specific process parameter is specifically instantiated as the spindle speed offset value of the target CNC machine tool or the set temperature offset value of the heating reactor. This adjustment range is typically set between 0.1% and 0.5% of the rated operating parameters. The physical basis for choosing this small perturbation range is that this level of perturbation is sufficient to induce an observable physical step change in the active power of the underlying asynchronous motor or the energy consumption of the heating tank, while ensuring that the various physical and chemical indicators of the end product remain within acceptable tolerances, thereby enabling targeted testing without compromising actual production yield.

[0047] S320, based on the aforementioned perturbation tags, the physical response verification unit sends a configuration update command to the edge computing gateway within a predetermined response time window after the procurement message reaches the target gateway, forcibly increasing the sampling frequency of the underlying physical sequence. The length of this predetermined response time window is specifically set based on a combination of the maximum round-trip time of the underlying industrial Ethernet and the mechanical transmission inertia time constant of the target device. Under normal steady-state conditions, the sensor sampling frequency is usually measured in minutes; however, during the perturbation response period, to capture the transient changes in physical energy consumption characteristics, this sampling frequency is dynamically increased to the second level. The physical response verification unit extracts the underlying physical sequence within the high-frequency sampling period and extracts its actual response gradient vector in the time domain through first-order difference operations. Simultaneously, the unit calls the locally stored production line digital twin process library and substitutes the perturbation amplitude into the equipment dynamics polynomial. This digital twin process library is built on system identification theory. It pre-fits the transfer function between equipment control commands and output physical parameters using the least squares method, thereby being able to deduce the theoretical state evolution trajectory of the system in the transient process based on the injected perturbation amplitude, and calculate and generate the corresponding expected physical response gradient vector accordingly.

[0048] S330, based on the extracted real and expected features, the physical response verification unit performs multi-dimensional vector space analysis to quantify the physical consistency between the two. The calculation process utilizes a normalized inner product algorithm with a smoothing factor to output a consistency score reflecting the strength and direction of the response. The specific calculation process for the response consistency score is shown in the formula: ; In the formula, Indicates the response consistency score; This represents the actual response gradient vector derived from the underlying real high-frequency sampled data; This represents the expected physical response gradient vector generated based on equipment dynamics calculations. This represents the dot product operation between vectors; This indicates solving for the magnitude of the second norm of the corresponding vector; This represents the function that takes the maximum value. This represents a preset minimum positive real number. The technical purpose of configuring this minimum positive real number is to effectively prevent the underlying program from crashing due to a zero denominator in division operations when the underlying physical system experiences a power outage or a malicious bypass of the physical gateway, resulting in a strictly zero magnitude of the actual response gradient. This minimum value is typically configured near the machine precision limit of the floating-point number.

[0049] In complex industrial environments, relying solely on the extreme values ​​of single vector inner products for judgment is susceptible to interference from occasional noise such as power grid transient surges or local sensor transmission interruptions. To avoid one-sided judgments, the physical response verification unit introduces joint judgment logic in the time and amplitude domains. Specifically, this unit further verifies the physical moment when the actual response gradient vector exhibits an abrupt edge exceeding the limit on the time axis, calculating the time difference between this and the moment the perturbation tag is issued. This unit verifies whether this time difference falls within the tolerance band allowed by the inherent physical time delay determined by the aforementioned benchmark self-learning stage. This tolerance band is specifically determined by floating above and below the inherent physical time delay by the maximum network jitter margin allowed by the system. Only when the time response conforms to the physical causality of the underlying process, and the response consistency score calculated in three consecutive detection cycles exceeds the set trust threshold, does the system determine that the underlying physical node is truly under control and has actually executed the production task. This trust threshold is strictly defined by fitting the probability density distribution of the scores of historical normal test samples and selecting the lower critical value of the 5% rejection region, with a typical value range between 0.85 and 0.90. For specific packet interception and parsing and industrial Ethernet packet reassembly operations, those skilled in the art can use deep packet inspection technology combined with standard protocol stack parsing. The underlying network data packet interaction mechanism is a well-known technology in this field and will not be elaborated upon here.

[0050] See attached document Figure 6 , Figure 6 This is a flowchart illustrating the weighted calculation of comprehensive early warning intensity and the mapping of abnormal behavior patterns according to an embodiment of the present invention. In this embodiment, the system deploys a comprehensive decision-making module in a cloud-based early warning center. This module performs multi-dimensional fusion of the independent indicators obtained from the preceding calculations, utilizes a nonlinear evaluation system to avoid misjudgments caused by occasional failures in a single testing stage, and maps the fusion results to specific business violation patterns to trigger corresponding control strategies.

[0051] S410, the integrated decision-making module obtains the phase deviation coefficient output by the phase deviation calculation unit and the response consistency score output by the physical response verification unit. To balance the risk tolerance of different orders, the integrated decision-making module simultaneously extracts the business sensitivity index of the current production order from the enterprise resource planning system. As a preferred method, this business sensitivity index is specifically obtained by linearly summing the order's cash flow amount and delivery urgency after performing maximum-minimum normalization. Considering the large differences in the original dimensions and numerical distribution of different order parameters, the integrated decision-making module normalizes these two parameters to the [0, 1] interval through this normalization operation. To prevent the division operation from overflowing and crashing during the normalization process due to the historical order range being strictly zero, such as multiple consecutive orders having completely identical cash amounts, the system forcibly superimposes a preset minimum smoothing constant when calculating the difference between the maximum and minimum values ​​in the denominator. The physical and commercial basis for selecting the above business parameters as input items is that the cash flow amount is directly related to the scale of economic losses caused by default or fraud, while the delivery urgency objectively characterizes the motivation intensity of supply chain nodes to take cheating measures due to insufficient capacity. Combining physical parameters with business parameters can make the final early warning judgment conform to actual business logic.

[0052] S420, based on the extracted multidimensional features, the comprehensive decision-making module introduces a nonlinear weighted algorithm to calculate the comprehensive early warning intensity. The general principle of this calculation process is to suppress the divergence of large-scale steady-state deviations through a logarithmic function, while simultaneously amplifying the penalty weight when physical probing fails through an inverse proportional function. This multidimensional nonlinear fusion logic effectively overcomes the deficiency of traditional linear weighting, which is easily diluted by a single normal indicator. The specific comprehensive early warning intensity calculation process is shown in the formula: ; In the formula, This indicates the overall early warning intensity score for the current production cycle; This represents the phase deviation coefficient of the output after operating condition modulation; This represents the response consistency score obtained through physical perturbation testing; This represents the business sensitivity indicators extracted from the business system; Represents a logarithmic function with the natural constant as its base; This represents the steady-state deviation weighting coefficient, and its value typically ranges from 0.3 to 0.5. This represents the dynamic trial penalty weighting coefficient, which typically ranges from 0.4 to 0.6. This represents the business sensitivity weighting coefficient, which typically ranges from 0.1 to 0.2. The specific value of the aforementioned weighting coefficient is determined using the analytic hierarchy process (AHP) combined with the historical expert scoring matrix, and satisfies the constraint that the sum of the three factors is a constant. This represents the preset smoothing penalty constant. The technical purpose of setting this smoothing penalty constant is to forcibly prevent the system overflow and crash caused by the division operation when the response consistency score decays to zero due to the complete blocking of probe commands by the underlying node, while ensuring that the warning intensity output is a very large finite warning value, which is typically set to 0.01 to 0.05.

[0053] S430, after obtaining the calculated comprehensive warning intensity, if it exceeds the system's preset static alarm limit, the comprehensive decision module activates the internally configured behavioral pattern mapping neural network to determine the specific physical violation type of the supply chain node. This behavioral pattern mapping neural network specifically adopts a multi-layer feedforward perceptron architecture. The internal hierarchical structure consists of an input layer with three neurons, three cascaded fully connected hidden layers, and an output layer with four neurons. Regarding data flow, the phase deviation coefficient, response consistency score, and comprehensive warning intensity are concatenated and integrated into a three-dimensional state feature vector, which is fed into the network via the input layer. To eliminate the absolute magnitude differences in features across dimensions due to their different physical origins, the comprehensive decision module pre-aligns the three-dimensional state feature vector using mean-variance standardization before feeding the data into the hidden layer. This prevents unprocessed extremely large values ​​from overwhelming the feature contributions of other dimensions and accelerates the gradient convergence process of the network's internal weights. Subsequently, the hidden layer uses a modified linear unit as the activation function to extract the nonlinear correlation mapping relationship in the multi-dimensional input sequence. The output layer uses a normalized exponential function for classification activation. The four dimensions of the output layer strictly correspond to four preset typical business physical states. Specific sub-characteristics include the compliant production state representing normal execution, the recorded playback attack state representing the alteration of historical data transmission, the fake state of equipment idling with the motor powered on but no materials being fed, and the illegal outsourcing state representing the subcontracting of orders to substandard factories.

[0054] S440, to ensure the completeness and accuracy of the inference of the above model, the system needs to perform supervised training on the behavior pattern mapping neural network in the offline stage. The comprehensive decision-making module collects multi-dimensional IoT logs of historical closed orders, and extracts three-dimensional state feature vectors as training samples through the same process described above. The classification labels of the samples are generated by auditors through cross-comparison and qualitative verification based on on-site video surveillance recordings, underlying industrial water, electricity, gas bills, and logistics inbound and outbound documents, and are processed through one-hot encoding. During the network backpropagation iteration process, the system uses the cross-entropy loss function to calculate the difference between the network's predicted probability distribution and the true labels. The specific loss function calculation process is shown in the formula: ; In the formula, This represents the cross-entropy loss value for a single training batch; This represents the preset total number of categories, which is 4 in this embodiment; Represents the th element in the true label vector. A Boolean value for each category, which takes the value of one when the sample actually belongs to that category, and the value of zero otherwise; The output layer of the behavior pattern mapping neural network is for the first... The probability values ​​predicted for each category; This represents the logarithmic function. This is to prevent the logarithmic function from being corrupted by the network outputting extremely zero probabilities early in training. The system generates a negative infinity numerical overflow error by pre-calculating the predicted probability when calculating the cross-entropy loss. A minimum truncation operation was performed, forcibly clamping its lower limit at... The numerical accuracy is improved. To avoid prediction bias caused by extreme imbalance in the number of samples in individual categories during the early stages of training, a dynamic learning rate decay mechanism is introduced simultaneously. After the feature vectors collected in real time are input into the trained network, the comprehensive decision module selects the pattern corresponding to the maximum value in the output probability distribution as the final qualitative conclusion of the anomaly, and encapsulates the associated underlying sequence and message interception record into a data storage package and sends it to the supply chain supervision platform. For specific business work order dispatch and data package storage on the blockchain, those skilled in the art can use smart contracts combined with distributed message queue technology for system-level integration. The underlying cross-domain system collaboration mechanism is a well-known technology in this field and will not be elaborated here.

[0055] Specific application examples are described below.

[0056] The target supply chain node is a manufacturer of ternary precursors for new energy batteries. The system deploys an edge computing gateway at its production site. Physical interfaces connect to industrial meters connected to the reactor agitator motor to obtain kinetic energy flow sequences, flow meters connected to the sedimentation process to obtain waste discharge sequences, and weighbridges connected to the plant area to obtain logistics throughput sequences. The system also obtains surface transaction sequences through an Enterprise Resource Planning (ERP) system application programming interface (API).

[0057] The edge gateway pre-samples and aligns the multi-source heterogeneous sequences to a 5-minute time step. The edge feature desensitization unit calculates the first derivative of the underlying physical sequence and applies a sign function to generate a phase feature stream containing only three discrete states: 1, 0, and -1. The reference matrix generation unit extracts data from the node's historical 90-day steady-state operation and uses a dynamic time warping algorithm to calculate the time offset between the physical energy consumption peak and the surface document weighting peak. The system selects the statistical median of the offsets corresponding to multiple optimal alignment paths, obtaining a standard physical time delay of 51.3 hours for this process, thereby constructing a standard reference phase matrix.

[0058] During the current monitoring period, the plant's ambient temperature decreased by 7.2 degrees Celsius compared to the historical steady-state average, and the total cumulative operating time of the core reactor increased by 4102 hours. The environmental equipment coupling modulation unit preprocessed these two state parameters and input them into the time-delay compensation neural network model. The model inference calculation yielded a time delay increment of 3.8 hours caused by the temperature drop and spindle wear. The environmental equipment coupling modulation unit performed matrix addition on this legitimate time delay increment and the standard reference phase matrix to generate a dynamic reference phase matrix with a time delay calibration of 55.1 hours. Subsequently, the phase deviation calculation unit used the time-domain cross-correlation function to find the current actual phase difference matrix and calculated the residual L2 norm between this matrix and the dynamic reference phase matrix, obtaining a phase deviation coefficient of 0.87.

[0059] A value of 0.87 triggered the system's first-level control limit. The message probe injection unit intercepted a routine electronic data interchange procurement message about to be sent to this node and wrote a perturbation tag in the extended field. This perturbation tag required the node to increase the operating frequency of the reactor agitator motor for the current batch by 0.15%. After the message reached the physical gateway, the physical response verification unit increased the sensor sampling frequency to 2 seconds within the response time window and extracted the actual response gradient vector from the underlying meter data. The system called the digital twin process library to generate the expected physical response gradient vector, performed an inner product operation including a smoothing factor on the two, and calculated a response consistency score of 0.08.

[0060] The early warning intensity calculation unit obtains a phase deviation coefficient of 0.87 and a response consistency score of 0.08, extracts the business sensitivity index for this batch of orders, configures nonlinear penalty weights for calculation, and outputs a comprehensive early warning intensity index. The abnormal mode execution unit inputs the above three-dimensional state feature vector into a multi-layer feedforward perceptron network. The classification result of the output layer of this behavior pattern mapping neural network points to the illegal outsourcing mode. The system finally outputs order freezing and backup supplier switching instructions to the downstream scheduling center through the enterprise service bus.

[0061] Table 1. Test Results of Multi-Dimensional Operational Indicators for Supply Chain Monitoring and Early Warning Algorithms

[0062] According to the data in Table 1, the monitoring and early warning algorithm of this invention exhibits definite numerical differences in various test indicators. In Table 1, the data record for the probe response calculation time indicator of the traditional fixed threshold comparison algorithm and the static sequence alignment cross-correlation algorithm is " / ", indicating that these two basic algorithms are not equipped with a perturbation probe targeting verification module and lack the functions of actively issuing probe commands and calculating underlying physical responses; therefore, this test data is missing. The traditional fixed threshold comparison algorithm relies solely on the absolute numerical limitations of business documents and underlying data, lacking a time series alignment mechanism. Its false alarm rate reaches 17.42%, and the average system confirmation delay time reaches 183.45 hours, making it unable to adapt to production rhythm fluctuations in complex industrial scenarios. The static sequence alignment cross-correlation algorithm introduces a data alignment mechanism, increasing the detection rate to 64.27% and shortening the average confirmation delay time to 82.11 hours. Because this algorithm does not include adaptive adjustment logic for external physical conditions, it is prone to judgment errors when facing environmental temperature fluctuations and natural equipment degradation; the 9.18% false alarm rate still causes unnecessary production interruptions.

[0063] The false alarm rate of the monitoring and early warning algorithm of this invention is reduced to 1.34%. Ambient temperature and depreciation parameters are extracted through an environmental equipment coupling modulation unit. The system uses a neural network to calculate the legal time delay increment and dynamically adjusts the judgment benchmark based on changes in the objective environment. This structure eliminates the interference of natural performance drift on the early warning mechanism, narrowing the algorithm's judgment limit to the purely human deviation range.

[0064] In terms of detection rate, the monitoring and early warning algorithm of this invention achieves 95.82%. For highly concealed violations such as data replay attacks that tamper with historical records and falsification of equipment idling energy consumption, the perturbation probe targeted verification module plays a physical verification role. The system forcibly verifies the transient response capability of the underlying physical production line by subtly attaching small-amplitude process parameter bias commands to the sent messages. The one-way data acquisition mode is transformed into a closed-loop feedback verification that includes action probing. The probe response calculation time remains at 1.14 seconds, reflecting the computational efficiency of collaborative operations between edge nodes and the cloud-based early warning center. Combined with multi-source phase feature comparison in the time dimension, the average system confirmation delay is compressed to 5.23 hours. The comprehensive decision module performs joint nonlinear weighting on phase deviation and response consistency scores, outputting a clear business physical state mapping result, enabling enterprises to have an operational window for implementing control interventions before problem batches are formed and delivered.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A new energy industry supply chain monitoring and early warning method based on behavior patterns, characterized in that, The method includes the following steps: S100: Obtain the underlying physical sequence and the surface transaction sequence, extract the first derivative to generate the current phase feature stream; align the historical phase feature stream with the historical surface transaction sequence, calculate the median time offset, and generate the standard reference phase matrix; S200 collects environmental state parameters and equipment depreciation state parameters, inputs them into a time delay compensation neural network model, calculates the legal time delay increment, adds it to the standard reference phase matrix to generate a dynamic reference phase matrix; calculates the current phase feature flow and the surface transaction sequence delay to construct the actual phase difference matrix; calculates the residual norm of the actual phase difference matrix and the dynamic reference phase matrix to solve for the phase deviation coefficient; When the phase deviation coefficient of S300 reaches the first-level control limit, the procurement message is injected with a perturbation tag and issued; the actual response gradient of the underlying physical sequence is collected, and the inner product of the actual response gradient and the expected response gradient is calculated to obtain the response consistency score. S400 calculates a comprehensive early warning intensity index by weighting the phase deviation coefficient and the response consistency score, determines abnormal behavior patterns, and outputs control instructions.

2. The new energy industry supply chain monitoring and early warning method based on behavioral patterns according to claim 1, characterized in that, In step S100, the underlying physical sequence includes a kinetic energy flow sequence representing the electrical load, a waste discharge sequence representing the sewage discharge situation, and a material throughput sequence representing the material inflow and outflow. The step of extracting the first derivative to generate the current phase feature flow specifically includes: The underlying physical sequence is aligned to a uniform discrete time step using an interpolation algorithm; The first derivative of the aligned underlying physical sequence in the time dimension is calculated, and the absolute amplitude parameter is discarded by applying a sign function to generate the current phase feature stream composed of discrete state parameters.

3. The new energy industry supply chain monitoring and early warning method based on behavioral patterns according to claim 1, characterized in that, In step S100, the step of aligning the historical phase feature stream with the historical surface transaction sequence, calculating the median time offset, and generating the standard reference phase matrix specifically includes: Construct a two-dimensional distance matrix and calculate the local Euclidean distance between the historical phase feature flow sampling points and the historical surface transaction sequence sampling points, and solve for the regular path with the minimum cumulative distance; Extract multiple sets of time offsets from at least five complete historical delivery cycles; The statistical median of multiple time offset sets is calculated as the inherent physical time delay, and a standard reference phase matrix is ​​constructed by assembling the inherent physical time delays.

4. The new energy industry supply chain monitoring and early warning method based on behavioral patterns according to claim 1, characterized in that, In step S200, the environmental state parameters include the plant ambient temperature and the power grid frequency deviation, and the equipment depreciation state parameters include the cumulative total operating time of the core production line and the spindle calibration wear degree; the specific steps of collecting the environmental state parameters and equipment depreciation state parameters and inputting them into the time delay compensation neural network model to calculate the legal time delay increment include: The environmental state parameters and equipment depreciation state parameters are concatenated and input into a time-delay compensation neural network model built on a multilayer perceptron architecture. The nonlinear cross-coupling effect contained in the parameters is extracted by a fully connected hidden layer, and the time delay elongation caused by temperature and wear is calculated by the output layer as a legitimate time delay increment.

5. The new energy industry supply chain monitoring and early warning method based on behavioral patterns according to claim 1, characterized in that, In step S200, the calculation of the residual norm of the actual phase difference matrix and the dynamic reference phase matrix to solve the phase deviation coefficient specifically includes: Calculate the difference matrix between the actual phase difference matrix and the dynamic reference phase matrix; Solve for the L2 norm contained in the difference matrix, divide the solved L2 norm by the sum of the L2 norm contained in the dynamic reference phase matrix and the preset smoothing positive real constant, and obtain the dimensionless phase deviation coefficient. When the incremental matrix and the standard reference phase matrix are misaligned due to missing sampling, zero-filling or dimensionality reduction truncation based on singular value decomposition is performed in advance to force the alignment of their dimensions.

6. The new energy industry supply chain monitoring and early warning method based on behavioral patterns according to claim 1, characterized in that, In step S300, the injection of perturbation tags into the procurement message specifically includes: When the phase deviation coefficient continuously exceeds the first-level control limit set based on three times the standard deviation of the historical residual mean within the set observation time window, the procurement message to be issued will be intercepted. A perturbation tag is dynamically written into the extended protocol field included in the procurement message. The perturbation tag is used to instruct the target node to make minor adjustments to specific process parameters of the current batch of raw materials. The adjustment range is set between 0.1% and 0.5% of the rated operating parameters.

7. The new energy industry supply chain monitoring and early warning method based on behavioral patterns according to claim 1, characterized in that, In step S300, the acquisition of the actual response gradient of the underlying physical sequence specifically includes: Within the predetermined response time window after the message reaches the target gateway, a configuration update command is sent to the edge computing gateway to force an increase in the sampling frequency of the underlying physical sequence; Extract the underlying physical sequence within the high-frequency sampling period, and extract the actual response gradient vector in the time domain through first-order difference operation; The production line digital twin process library is invoked to substitute the perturbation amplitude into the equipment dynamics polynomial to generate the expected response gradient.

8. The new energy industry supply chain monitoring and early warning method based on behavioral patterns according to claim 1, characterized in that, In step S400, the weighted average of the phase deviation coefficient and the response consistency score to obtain the comprehensive early warning intensity index specifically includes: Extract the current production order business sensitivity index from the enterprise resource planning system. The business sensitivity index is obtained by linearly summing the order cash flow amount and the delivery urgency after performing maximum and minimum normalization. The steady-state deviation penalty term included in the phase deviation coefficient is calculated using a logarithmic function, and the dynamic trial penalty term included in the response consistency score is calculated using an inverse proportional function. A comprehensive early warning intensity index is generated by performing a linear combination weighted average of the steady-state deviation penalty term, the dynamic trial penalty term, and the business sensitivity index.

9. The new energy industry supply chain monitoring and early warning method based on behavioral patterns according to claim 1, characterized in that, In step S400, determining the abnormal behavior pattern specifically includes: The phase deviation coefficient, response consistency score, and comprehensive early warning intensity index are concatenated and integrated into a three-dimensional state feature vector. The mean-variance standardization operation is used to perform dimensional alignment on the three-dimensional state feature vectors. The aligned 3D state feature vectors are fed into a behavior pattern mapping neural network containing a multi-layer feedforward perceptron architecture, and the classification probability distribution is output through a normalized exponential function. The pattern corresponding to the maximum value in the classification probability distribution is selected as the final qualitative conclusion of the anomaly, and it is classified as a behavioral anomaly pattern.

10. A new energy industry supply chain monitoring and early warning system based on behavioral patterns, characterized in that, The system using the new energy industry supply chain monitoring and early warning method based on behavioral patterns as described in any one of claims 1-9 includes: The multidimensional feature perception and benchmark self-learning module includes an edge feature desensitization unit and a benchmark matrix generation unit. The edge feature desensitization unit is used to obtain the underlying physical and surface transaction sequences and extract the first derivative to generate the current phase feature stream. The benchmark matrix generation unit is used to align the historical sequences, calculate the median time offset, and generate a standard benchmark phase matrix. The adaptive modulation and deviation calculation module includes an environmental equipment coupling modulation unit and a phase deviation calculation unit. The environmental equipment coupling modulation unit is used to calculate the legal time delay increment based on environmental and equipment depreciation status parameters and generate a dynamic reference phase matrix. The phase deviation calculation unit is used to calculate the phase deviation coefficient by calculating the residual norm of the current actual phase difference matrix and the dynamic reference phase matrix. The perturbation probe targeting verification module includes a message probe injection unit and a physical response verification unit. The message probe injection unit is used to inject perturbation tags into the procurement message when the phase deviation coefficient reaches the first level control limit. The physical response verification unit is used to collect the underlying actual response gradient and calculate the inner product with the expected response gradient to obtain the response consistency score. The integrated decision-making and control execution module includes an early warning intensity calculation unit and an abnormal mode execution unit. The early warning intensity calculation unit is used to calculate the comprehensive early warning intensity index by weighting the deviation coefficient and the response consistency score. The abnormal mode execution unit is used to determine the abnormal behavior mode and output control instructions.

Images