A method of plasticizing order matching

Abstract

The present application relates to the technical field of intelligent manufacturing, and more particularly to a plasticizing order matching method, which comprises the following steps: first, encoding the order and production line data events, generating a shape vector and an energy flow vector through a neural network; constructing a coupling matrix in a hyperbolic space and matching once using an Ising model; if the output gradient is over the threshold, refining through neural morphology inhibition and calculating the local entropy gradient; feeding the refined results and production line feedback into a causal generation graph to determine conflicts, and if necessary, modulating a solver to generate a correction matching; finally, writing the matching vector and conflict markers into a blockchain, and each node updates the model through federated learning, and drives the curvature and threshold adaptively through the shape entropy. The method realizes real-time production scheduling with low energy consumption, less line change and sustainable learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing, and particularly relates to a plasticization order matching method. BACKGROUND

[0002] In the core link of production scheduling in the continuous extrusion industry of polyolefin, modified engineering plastics and the like, orders of different formulas, colors and melt indexes will produce cross residues and energy consumption climbing under the joint action of screw and cylinder, which directly affects the unit energy consumption, line changing frequency and product consistency; the higher the production scheduling efficiency is, the higher the production line utilization rate and delivery rate are, and the more significant the overall supply chain is increased.

[0003] The existing scheme mostly adopts rule scheduling or integer programming: the rule scheduling arranges orders in sequence after grouping them according to brand and color, and ignores the thermal inertia of the production line, resulting in high energy consumption and frequent machine cleaning; the integer programming is based on Euclidean distance or linear weight evaluation, and it is difficult to capture high-dimensional dynamic characteristics, and the model re-computation is time-consuming, and it cannot respond to sudden orders online; the traditional optimization lacks closed-loop learning, and the scheduling strategy is fixed, and cannot be updated according to the wear and tear of the production line and the energy consumption curve, resulting in energy waste, frequent line changing and rigid scheduling. SUMMARY

[0004] In view of the problems existing in the prior art, the present application provides a plasticization order matching method, which maps orders and production line signals into pulse sequences by event coding, and outputs morphological vectors and energy flow vectors through a neural network; a coupling matrix is constructed in hyperbolic space and a low-energy match is obtained from an Ising model; a neural morphological inhibition smooths the energy gradient in real time, and a causal generation graph combines with physical feedback to detect conflicts and correct deviations; the matching result is written into a chain, the federal learning updates the full-stack parameters, the cumulative morphological entropy drives the curvature self-adjustment, and finally an energy-efficient, low-line-changing and sustainable learning scheduling system is realized.

[0005] A plasticization order matching method, comprising the following steps:

[0006] Event coding is performed on the order parameter stream and the production line parameter stream, and a neural network is used to generate order morphological vectors and production line energy flow vectors, and an initial energy flow potential field is formed;

[0007] A coupling matrix is constructed according to the distance between the morphological vectors and the energy flow vectors in a non-Euclidean space, the coupling matrix is input into an Ising model solver to generate a first matching result; when the signal gradient output by the Ising model solver exceeds a preset threshold, an inhibition signal generated by a neural morphological calculation model is used to modulate the Ising model solver, a refined matching result is generated and a local entropy gradient is calculated;

[0008] The refined matching result, local entropy gradient and production line physical feedback are input into the causal generation graph model to calculate the conflict probability; when the production line state change rate and the conflict probability both exceed the preset threshold, the Ising model solver is modulated again to generate the correction matching result and output the conflict label;

[0009] The morphological vector, energy flow vector and conflict label corresponding to the correction matching result are written into the distributed ledger, and each node updates the neural network parameters, Ising model solver parameters, neural morphological calculation model parameters and causal generation graph model parameters through contrastive learning and federated aggregation; when the cumulative morphological entropy decrease rate is lower than the preset threshold, the non-Euclidean space curvature, signal gradient threshold and inhibition signal gain are adjusted, and the parameters are updated synchronously.

[0010] Preferably, the event code measures the adjacent time difference of the continuous sampling signal in a fixed length sliding window, and converts the time difference into a pulse sequence.

[0011] Preferably, the neural network sequentially sets a time convolution layer, a gated linear layer and a global average pooling layer, and the output order morphological vector dimension is greater than the production line energy flow vector dimension.

[0012] Preferably, each element of the coupling matrix is generated by substituting the distance between the morphological vector and the energy flow vector in the hyperbolic space into the exponential mapping function, and the exponential coefficient of the exponential mapping function is adjusted in real time according to the hyperbolic space curvature.

[0013] Preferably, the output signal gradient of the Ising model solver is obtained by performing a weighted average operation on the continuous light intensity sampling values, and the weighted coefficient is updated adaptively with the light intensity change rate.

[0014] Preferably, when the output signal gradient of the Ising model solver exceeds the preset threshold, the neural morphological calculation model reduces the corresponding synaptic weight to generate an inhibition signal, and uses the inhibition signal to modulate the Ising model solver.

[0015] Preferably, the local entropy gradient is obtained by multiplying the Euclidean distance, energy consumption prediction value and switching cost by preset weights and then summing them.

[0016] Preferably, the causal generation graph model generates counterfactual order scenarios using diffusion backtracking, and calculates the conflict probability according to the number of times of mismatching of the counterfactual order scenarios.

[0017] Preferably, the distributed ledger uses a Byzantine fault-tolerant consensus algorithm to record the hash values corresponding to the morphological vector, energy flow vector and conflict label, and adds a timestamp to each record.

[0018] Preferably, each node performs federated averaging algorithm aggregation to update neural network parameters, Ising model solver parameters, neuromorphic computing model parameters and causal generative graph model parameters according to the proportion of the number of local samples, and broadcasts the updated parameters to all nodes.

[0019] Compared with the prior art, the application has the advantages and beneficial effects that:

[0020] Through the event coding + pulse neural network technical means, the dimensional unification and sparse compression of the multi-dimensional dynamic characteristics of orders and production lines are realized; through the hyperbolic distance coupling matrix + Ising solution technical means, the millisecond-level global minimum energy matching effect is realized; through the neuromorphic inhibition technical means, the sub-millisecond adaptive refinement and energy consumption surge suppression effect is realized; through the causal generative graph + physical feedback technical means, the conflict probability accurate evaluation and online correction effect is realized; through the distributed ledger + federated aggregation technical means, the model parameter credible synchronization and long-term self-learning effect are realized. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The figure is a flowchart of the method of the application;

[0022] Figure 2 The figure is a schematic diagram of the distributed ledger and federated update in the application. DETAILED DESCRIPTION

[0023] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the present disclosure. However, it will be apparent to one of ordinary skill in the art that one or more embodiments can be practiced without these specific details. In addition, in the following description, descriptions of well-known structures and techniques have been omitted to avoid unnecessarily obscuring the concept of the present disclosure.

[0024] The terms used herein are merely used to describe specific embodiments and are not intended to limit the present disclosure. The terms "include", "comprise" and the like used herein indicate the presence of the described features, steps, operations and / or components, but do not exclude the presence or addition of one or more other features, steps, operations or components.

[0025] All terms used herein (including technical and scientific terms) have meanings commonly understood by one of ordinary skill in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted to have meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0026] As Figure 1 shown, a plasticized order matching method comprises the following steps:

[0027] The order parameter flow and the production line parameter flow are event coded, and an order form vector and a production line energy flow vector are generated through a neural network to form an initial energy flow potential field;

[0028] The process of event coding the order parameter flow and the production line parameter flow, generating the order form vector and the production line energy flow vector through the neural network, and forming the initial energy flow potential field is the "data embryo" of the subsequent matching link of the application. First, the order parameter flow can be regarded as a multi-dimensional discrete sequence of orders arriving over time, containing information such as material brand, color ratio, target melt index, delivery period priority, etc.; the production line parameter flow comes from real-time sampling of screw torque, back pressure, barrel temperature distribution, material level feedback, etc. If the original numerical value or low-frequency average value is directly input into the solver, it is easy to cause distance measurement distortion due to dimensional differences and sudden jumps. Therefore, the application maps continuous sampling to a pulse time difference sequence through event coding. The specific method is to record the time interval of adjacent sampling points within a fixed sampling window, represent the forward transition as a pulse, the shorter the time difference, the denser the pulse count, and the reverse or no transition does not generate a pulse. This encodes the amplitude information into the pulse density implicitly, which reduces the bandwidth and retains the dynamic characteristics.

[0029] The pulse sequence after event coding is input into the neural network for vectorization. The neural network is composed of a time series convolution layer, a gated linear layer and a global average pooling layer. The design logic is to first extract the local mode of pulse density in the convolution layer, then filter noise with the gated linear layer, and finally form a fixed-length vector through pooling. The convolution kernel size is sufficient to cover a single process cycle as the principle, and the gated linear layer uses the product of the gating coefficient and the pulse intensity to suppress abnormal pulses. After training, the high-dimensional form vector output at the order end can highlight the formula similarity, and the energy flow vector output at the production line end can depict the production line thermal inertia, residual material spectrum and other working conditions.

[0030] In order to make the matching process have energy constraints, the application generates an initial energy flow potential field by weighted integration of the energy flow vector. The potential field can be summarized as follows:

[0031]

[0032] Where U j is the potential density of the jth production line, ψ j,k (t) represents the kth component of the energy flow vector, w k is the dimension weight, and T is the integration window length. The weight is distributed according to the historical energy consumption contribution. The role of the potential field is similar to the "potential energy" of the continuous working condition. When constructing the coupling matrix, the subsequent Ising model solver multiplies and U j , so that the production line with low potential energy naturally has priority in matching.

[0033] Example: A batch of high melt index polypropylene orders and a batch of low melt index orders were simultaneously reported. After event encoding, the pulse density of the high melt index orders showed a concentrated peak, which the neural network mapped to the same sub-region in the morphological vector space; while production line A, due to its smaller screw wear coefficient and lower temperature fluctuation, had the shortest distance to the high melt index orders in the energy flow vector space, coupled with the potential energy density U A The value is relatively low, so it is preferentially selected by the Ising model solver. At this point, even if another production line B also meets the process conditions, due to U... B A higher weight will result in a smaller weight in the coupling matrix, thus avoiding increased energy consumption. Through this mechanism, the present invention can reduce unit energy consumption and reduce material replacement residue without adding new hardware.

[0034] Preferably, the event encoding measures the adjacent time difference of continuously sampled signals within a sliding window of fixed duration and converts the time difference into a pulse sequence.

[0035] Event encoding is a key component of this invention, used to unify the dimensions of order parameter streams and production line parameter streams, compress bandwidth, and highlight dynamic information. The order parameter stream consists of data arriving over time, such as material grade, additive ratio, target melt index, and delivery priority; the production line parameter stream comes from real-time sampling of screw torque, back pressure, barrel temperature field, and material level. These two types of signals differ significantly in both dimensions and update rates. Directly feeding them into the subsequent vectorization network would lead to an unbalanced feature dominance effect, thus affecting matching accuracy. This invention introduces event encoding at the acquisition end, converting "signal amplitude" into a unified representation of "pulse time difference."

[0036] The principle of event coding can be summarized as follows: within a sliding window of length Δt, the signal s(t) is continuously sampled, and adjacent sampling times t are recorded. k With t k+1 The time difference τ between k =t k+1 -t k If τ k Less than the threshold τ min This indicates that the signal rises rapidly, triggering a "positive pulse"; if τ k Greater than the threshold τ max τ indicates that the signal changes or decreases slowly and does not trigger a pulse; k A "suppression pulse" is triggered when the pulse intensity is between the two thresholds. To prevent the pulse density from increasing indefinitely with the absolute amplitude, this invention uses an exponential suppression function to limit the maximum firing rate. Pulse intensity p k Calculated by the following formula:

[0037] When τ k ≤τ min

[0038] p k = 0 when τ k ≥ τ max

[0039] where λ is a time constant for controlling the sensitivity to short-time dramatic changes. The shorter the time difference, the higher the pulse intensity; beyond τ max it is considered as no pulse. Through this mapping, the amplitude and slope of the original signal are converted into the pulse firing rate and intensity, so that different source signals have comparability before entering the neural network.

[0040] In actual implementation, the sliding window Δt depends on the process beat. Example embodiment: the screw torque update period on a certain production line is about 20 milliseconds, and the temperature field refresh period is about 200 milliseconds. Setting Δt as 200 milliseconds can cover the slowest signal while not distorting the details of the fastest signal. The event encoder generates high-frequency positive pulses for the torque stream and sparse pulses for the temperature stream. After entering the same pulse convolution network, the convolution kernel can automatically learn the weight without manual normalization.

[0041] The effects of event encoding are reflected in the following aspects: when the original sampling rate is 10 kHz, 10,000 floating-point numbers per second need to be transmitted; after event encoding, the average pulse count is reduced to 2,000 times, and the bandwidth is reduced by 80%, leaving bandwidth for subsequent high-frequency writing of the distributed ledger. Traditional discrete sampling lacks amplification mechanism for sudden fluctuations, and event encoding emphasizes short-time dramatic changes through exponential inhibition function, so that abnormalities such as screw abnormal sound and material level sudden drop are actively pulled away in the shape vector space, and the matching engine can avoid risky production lines in time. Whether it is kilopascal scale of back pressure or percentage of formula proportion, it is ultimately in the form of pulse time difference into the network, and the distance measurement is no longer disturbed by the difference in physical dimensions.

[0042] In the matching process from order to production line, the calculation quality of shape vector and energy flow vector determines the accuracy of hyperbolic distance. Through event encoding, long sequence signals are compressed into pulse sequences, and neural networks only need to focus on the sparse patterns of pulses. The convolution layer extracts local resonance, the gate linear layer filters false pulses (such as slow fluctuations caused by inertia of temperature control system), and the global average pooling outputs fixed-length vectors. The shape vector dimension output by the order side is set to 128, and the energy flow vector dimension output by the production line side is set to 64. In the hyperbolic space, the shape vector dimension is higher than the energy flow vector dimension, which meets the design principle that the "demand space" is larger than the "capability space".

[0043] Further example: Order A requires high melt index polypropylene with low torque fluctuation tolerance; production line P has a dense torque pulse sequence and a sparse temperature pulse sequence, indicating that the equipment is in a stable high-torque zone. The hyperbolic space distance between the shape vector and the energy flow vector is denoted as d APIf the pulse intensity is not event-encoded, the high-sensitive features of order A will be diluted by the torque mean, resulting in d AP being too small, leading to matching errors. After event-encoding, the short-time torque peak triggers multiple positive pulses, and the "sensitive" dimensions of the morphology vector are highlighted, d AP increases, and the matching is more in line with the process safety requirements.

[0044] τ k represents the sampling time difference, in milliseconds; λ is the time constant; p k is the pulse intensity, dimensionless. For the same signal, a uniform λ is used for all sliding windows to ensure time domain consistency; different signals can set different τ min and τ max to adapt to the dynamic range of the signal. In the embodiments of the present application, torque τ min = 2 milliseconds, τ max = 10 milliseconds, temperature τ min = 20 milliseconds, and τ max = 100 milliseconds. It has been experimentally verified that the pulse count is stable and sensitive to process disturbances.

[0045] Preferably, the neural network is sequentially provided with a time sequence convolution layer, a gated linear layer, and a global average pooling layer, and the output order morphology vector dimension is greater than the production line energy flow vector dimension.

[0046] The neural network in the present application plays the role of "high-dimensional feature generator", and its output directly determines the accuracy of the subsequent coupling matrix and the convergence speed of the Ising model solver. The network is divided into three sections from input to output: time sequence convolution layer, gated linear layer, and global average pooling layer.

[0047] The core task of the time sequence convolution layer is to extract the local resonance pattern in the pulse sequence. The convolution kernel length is the same as the event-encoding window, so that one convolution can completely cover one process beat. In implementation, the convolution kernel weight is learned through backpropagation, and no preset filtering function is used. This end-to-end training method can automatically discover discriminative substructures, such as the complex pattern "torque surge-temperature slightly rise-back pressure stable" which is difficult to describe with threshold rules. The mathematical expression of the convolution operation is written as:

[0048]

[0049] where o n represents the nth sample of the convolution output sequence, x n-i+1 is the pulse input, the ith dimension weight w i is obtained through training, and b is the bias term. The convolution output is linearly rectified and then enters the next layer. L is the length of the convolution kernel.

[0050] The gated linear layer employs a gated linear unit (GLU) structure to suppress noise impulses. Each GLU consists of two parallel fully connected paths: one outputs the main signal, and the other outputs the gating coefficients. The two paths are multiplied element-wise to obtain the gated activation value. GLUs are chosen because pulse sequences can contain abrupt pseudo-pulses, such as switching jitter and sampling glitches. By learning the gating coefficients, the GLU automatically filters out such invalid information during the inference phase, improving the robustness of the vector representation. The gating operation can be written as:

[0051] y=σ(W g x+b g )⊙(W m x+b m )

[0052] Where y is the gated layer output, σ is the logic function, ⊙ is element-wise multiplication, and W g With W m These are the gate weights and the sovereign weights, respectively. In the formula, x is the output feature vector of the convolutional layer. Through this mechanism, impulses that truly contain process characteristics are preserved, while random noise is suppressed.

[0053] The global average pooling layer averages the gated features over time, outputting a fixed-length vector. The order-side output is called the order form vector, with a dimension of 128; the production line-side output is called the production line energy flow vector, with a dimension of 64. Dimension design follows the principle of "demand dimension greater than capacity dimension": the order form vector needs to accommodate high-dimensional information such as formula, physical properties, and delivery priority; the production line energy flow vector focuses on operating conditions such as energy consumption and residual material spectra, and its dimension can be appropriately reduced to alleviate subsequent computational load. Through dimensional mismatch, the form vector distribution in hyperbolic space is more dispersed, which is beneficial for forming a distinguishable distance structure with the energy flow vector.

[0054] In the initial potential field formation stage, the system performs a weighted integral of the production line energy flow vector to quantify the production line load state. Let the production line energy flow vector be ψ(t), the weight be w, the integration window length be T, and the potential energy density be calculated using the formula:

[0055]

[0056] In this formula, U reflects the average energy consumption tendency of the production line within the window, and the diagonal of the subsequent coupling matrix will be multiplied by e. -U Energy consumption penalties are implemented to prevent high-energy-consuming production lines from being assigned to low-priority orders.

[0057] The effectiveness of this network structure was verified in three aspects. First, the mean squared error was compared on a simulation dataset: the event encoding plus convolutional network scheme reduced the error by 30% compared to the pure mean pooling scheme. Second, online production line experiments were conducted: the number of line changes for the same batch of 80-ton polypropylene orders decreased from 12 to 7 under different network settings, while the unit energy consumption decreased by 5%. Third, abnormal order identification: after introducing illegal brand orders, the gated linear unit reduced the gate coefficient to below 0.3, triggering manual review by the system to avoid invalid production scheduling.

[0058] Example: The order flow contains 40 orders each of high melt index polypropylene and low melt index polyethylene, randomly mixed and arriving. The system first performs event encoding to obtain a pulse sequence; a convolutional layer captures the high pulse frequency of high melt index orders; a gating layer suppresses pseudo-pulses generated by temperature jitter; and global average pooling outputs a 128-dimensional shape vector. Torque, back pressure, and temperature at the production line end are processed by the same network to output a 64-dimensional energy flow vector. The system calculates the distance between the shape vector and the energy flow vector in hyperbolic space to construct a coupling matrix. The final matching results show that high melt index orders are preferentially matched to production lines with large torque margins, while low melt index orders flow into production lines with lower loads, achieving energy balance.

[0059] In accordance with the overall process of this invention, the three-layer structure of the neural network corresponds to three functions: convolutional layers for mining local temporal patterns, gated linear layers for noise reduction, and pooling layers for uniform length. Compared with traditional static feature concatenation schemes, this method eliminates the need for manually setting dimensional scaling factors, as everything is learned by the network, reducing human error. The convolutional structure supports parallel processing of multiple pulse sequences, keeping computational latency in the millisecond range. The sparse input introduced by event encoding further reduces the number of network parameters, allowing the network to be deployed on edge inference devices, providing margin for subsequent hyperbolic space operations and Ising model solvers.

[0060] In non-Euclidean space, a coupling matrix is ​​constructed based on the distance between the morphological vector and the energy flow vector. The coupling matrix is ​​then input into the Ising model solver to generate the first matching result. When the gradient of the output signal of the Ising model solver exceeds a preset threshold, the inhibition signal generated by the neuromorphic computing model is used to modulate the Ising model solver to generate a refined matching result and calculate the local entropy gradient.

[0061] This invention employs a scheme that couples a non-Euclidean distance metric with an Ising model solver during the plasticizing order matching process. Its core idea is to amplify the differences in order features within the hyperbolic geometric domain, and then achieve global matching through low-energy spin configuration.

[0062] The shape vector and energy flow vector describe the demand characteristics at the order end and the capacity characteristics at the production line end, respectively. Due to the exponential expansion characteristic of hyperbolic space with negative curvature, the same Euclidean interval will exhibit a larger distance difference in hyperbolic metric, thus amplifying fine-grained differences. Let the order shape vector be φ.i The production line energy flow vector is ψ j In the hyperbolic sphere model, the distance d between two vectors ij Formula used:

[0063]

[0064] Where ||·|| represents the Euclidean norm. To ensure numerical stability, the network output vector is normalized so that it falls within the open sphere. Using distance d... ij Constructing the coupling matrix J ij Defined as:

[0065] J ij =exp(-γd ij )·exp(-U j )

[0066] γ is the amplification factor, U j Let J be the potential energy density of the j-th production line within the integration window, used to penalize high-energy-consuming production lines. Coupling matrix J ij A larger value indicates a stronger coupling between order i and production line j, and the Ising model solver will tend to match the two. After the coupling matrix is ​​written into the solver, the spin variable σ... i With σ j Through the Hamiltonian:

[0067]

[0068] Energy minimization is performed. The solver locks into a low-energy configuration within a few microseconds, obtaining the first matching result.

[0069] To improve search stability, this invention introduces a neuromorphic inhibition mechanism. When the solver output signal gradient is too high, it indicates that the matching graph still has an energy slope, making it prone to mismatches. The neuromorphic computation model monitors the output signal gradient. like Where η is a preset threshold, the synaptic weights in the same direction as the slope are dynamically reduced to generate an inhibition signal. α is the suppression coefficient. After receiving the suppression signal, the solver re-mode-locks and converges to the refined matching result in just a few hundred nanoseconds. This step reduces the jump caused by the sharp drop in energy and ensures that subsequent local adjustments are differentiable.

[0070] After refining the matching results, the local entropy gradient G needs to be calculated to assess the degree of local stability. ij This invention weights distance, energy consumption, and production line switching costs into a single factor:

[0071] G ij =β1d ij +β2E ij +β3C ij

[0072] β1, β2, and β3 are the distance weight, energy consumption weight, and switching cost weight, respectively, and E ij C represents the predicted unit energy consumption of order i on production line j. ij This represents the cleaning and material change costs required to switch the order to production line j. Matching pairs with high local entropy gradients are given special attention in the subsequent causal correction process.

[0073] Example: A 120-unit monoolefin order and 5 twin-screw extrusion lines were used. First, the coupling matrix was calculated using the formula above and input into the Ising model solver. After 5 microseconds, the solver generated the matching results, and the output signal gradient norm was 0.8. The threshold η was set to 0.5, triggering an inhibition mechanism. The neuromorphic computation model generated an inhibition signal, reducing the gradient direction terms in the coupling matrix by an average of 15%. After the solver re-mode-locked for 1 microsecond, the signal gradient decreased to 0.3, and the refined matching results were output. The local entropy gradient G was calculated for each matching pair. ij The distribution ranges from 0.05 to 0.22. Matching pairs with a value higher than 0.18 were fed into a causal generative graph model for counterfactual testing, identifying three potential conflicts due to high delivery priority. After further optimization of the solver, the final near-completion rate improved by 3%. Simultaneously, the total energy consumption of the production line was reduced by 4% compared to traditional Euclidean distance matching.

[0074] Preferably, each element of the coupling matrix is ​​generated by substituting the distance between the morphological vector and the energy flow vector in hyperbolic space into an exponential mapping function, the exponential coefficient of which is adjusted in real time according to the curvature of the hyperbolic space.

[0075] The core of this invention lies in constructing a coupling matrix using non-Euclidean distance, then using the Ising model solver to complete a global matching of orders and production lines, and achieving sub-millisecond refinement through a neuromorphic inhibition mechanism.

[0076] The non-Euclidean distance is chosen from hyperbolic space distance because hyperbolic space has an exponential expansion characteristic due to its negative curvature: when two points are equidistant in Euclidean space, the two points in hyperbolic space will exhibit greater geometric separation due to curvature. In the plastics order matching scenario, even a small difference between the order shape vector and the production line energy flow vector needs to be amplified to avoid mismatch; therefore, hyperbolic distance is superior to Euclidean distance. Let the order shape vector be φ. i The production line energy flow vector is ψ j Both are first normalized and mapped to the interior of the open sphere of the hyperbola model, and then the distance is calculated:

[0077]

[0078] Where ||·|| represents the Euclidean norm, and arccosh is the inverse hyperbolic cosine function. To map distance to Ising coupling strength, this invention designs an exponential mapping function:

[0079] J ij =exp(-γd ij )·exp(-U j )

[0080] In the formula J ij These are the elements of the coupling matrix; γ is the exponential coefficient; U j This represents the production line potential energy density, used to penalize high-energy-consuming production lines. Unlike existing methods that fix γ, this invention adjusts the exponential coefficient in real time based on the hyperbolic curvature κ, with the following relationship: The more negative the curvature value, the stronger the spatial expansion, requiring a larger exponential decay coefficient to keep the coupling strength within a controllable range and avoid gradient explosion. The curvature parameter is derived from the entropy modulation result of the previous cycle, thus achieving adaptive geometric tuning.

[0081] After the coupling matrix is ​​written into the Ising model solver, the solver uses the spin variable σ i With σ j Minimize Hamiltonian:

[0082]

[0083] Positive spin indicates a match, and negative spin indicates a mismatch. The solver can use either optical coherence or electronic quantum annealing. This invention is not limited to a specific physical implementation, only requiring the output of the first match result within microseconds.

[0084] If, after the initial output of the Ising model solver, the gradient of the output signal is found to be too large, indicating a local steep slope in the matching, stability will be difficult to guarantee. The signal gradient is defined as... when When the threshold η is exceeded, the neuromorphic computational model enters the active state. The neuromorphic computational model monitors the solver output channel, reduces the corresponding coupling weights through synaptic plasticity, and generates an inhibition signal ΔJ. ij Specifically:

[0085]

[0086] Where α is the suppression coefficient. The solver accepts ΔJ. ij A fast annealing process is then performed, outputting a refined matching result. At this point, the local steep slopes in the matching graph are smoothed out, and the spin configuration is stabilized.

[0087] The refined matching results are used to calculate the local entropy gradient, which measures the robustness of the matching pair. Local entropy gradient G ij The Euclidean distance D between the order and the production line ijUnit energy consumption prediction E ij And switching costs C ij The weighted average of the three factors yields:

[0088] G ij =β1D ij +β2E ij +β3C ij

[0089] The three weights β1, β2, and β3 satisfy β1 + β2 + β3 = 1, and are determined offline by multi-objective optimization before deployment. Matching pairs with large local entropy gradients indicate the existence of potential instability factors, which will be counterfactually tested by the subsequent causal generative graph model.

[0090] In this example, the experimental factory has production lines P1 to P5, with an order batch size of 120 units. It is initialized with a hyperbolic curvature k = -1 and an exponential coefficient γ = 1. The initial solution yields... If the value exceeds the threshold η = 0.5, inhibition is triggered. The neuromorphic computational model outputs ΔJ. ij On average, the corresponding edge weights are reduced by 15%, and the solver completes the second annealing in just 0.5 microseconds. The value was reduced to 0.28. The local entropy gradient distribution was refined based on the matching results, with a maximum value of 0.22 corresponding to three order matching pairs. A causal generative graph model generated 20 counterfactual scenarios for these three matching pairs, identifying two pairs with conflict risks. Therefore, the coupling matrix was fine-tuned again, bidirectionally swapping conflicting orders, reducing the maximum local entropy gradient to 0.16. The final matching scheme resulted in no unplanned line changes during the subsequent 8 hours of production, a 4% reduction in unit energy consumption compared to the Euclidean distance baseline scheme, and a 41.7% reduction in the number of line changes.

[0091] This invention amplifies differences using hyperbolic distance, harmonizes coupling strength using exponential mapping, and balances convexity using curvature adaptive balancing. It is further supplemented by neuromorphic inhibition mechanisms and local entropy gradient evaluation, enabling the matching process to converge rapidly while maintaining globally optimal energy consumption and stability. Compared to traditional linear weighted or static Euclidean distance methods, this scheme improves the quasi-intersection rate by an average of approximately 3% and reduces the number of rearrangements by approximately 40% in multiple batches of comparative experiments, fully demonstrating the advantages of combining non-Euclidean coupling with adaptive inhibition.

[0092] Preferably, the gradient of the output signal of the Ising model solver is obtained by performing a weighted average operation on continuous light intensity samples, and the weighting coefficients are adaptively updated with the rate of change of light intensity.

[0093] In this invention, the Ising model solver can be implemented using optical coherence or charge-coupled methods. Regardless of the carrier form, its operating state is externalized as a time-varying output signal. This embodiment uses an optical coherence Ising solver as the subject of this study, and the output signal is embodied in the continuous sampling I(t) of light intensity over time t. To evaluate whether the solver is currently in a steep energy region, the light intensity values ​​at several moments need to be converted into a gradient index. If the gradient index is too large, it indicates that the system is not yet stable, and a neuromorphic suppression mechanism is required. This invention proposes the concept of a "weighted average gradient," which uses a continuous light intensity sampling sequence with adaptive weights to obtain the gradient. This suppresses detection noise while preserving rapid fluctuation information.

[0094] Suppose the solver outputs N light intensity sample values ​​I1, I2, ..., I during one mode-locking cycle. N The sampling interval is fixed at δt. This invention defines the discrete gradient of light intensity ΔI. k =|I k+1 -I k | Then use weight w k Perform a weighted average on it:

[0095]

[0096] In the formula Let ΔI be the gradient of the output signal; k w represents the light intensity increment within the k-th sampling interval. k For the corresponding weights.

[0097] The weight design follows the principle of "the faster the change, the greater the weight," to highlight rapid, short-term changes. The weight update rule adopts an exponential adaptive formula:

[0098]

[0099] in Let β represent the rate of change of light intensity within the k-th sampling segment, and β be the sensitivity coefficient. When the light intensity changes drastically... As the light intensity increases, the weights are exponentially amplified and their proportion increases in the subsequent averaging process; if the light intensity changes gradually, the weights naturally decay, thus reducing the contribution of slowly varying noise to the gradient. k Let be the k-th light intensity sample value, δt be the sampling interval, and β be a positive real number.

[0100] This weighted average gradient has two key advantages. First, by suppressing low-frequency noise through weighting, it avoids misinterpreting the slow drift during device warm-up as an energy ramp. Second, through a rate-driven exponential amplification mechanism, it maintains high sensitivity to high-frequency signals, enabling it to capture energy drops caused by spin flips within microseconds and promptly trigger neuromorphic inhibition.

[0101] In terms of implementation details, light intensity sampling is performed by a multi-channel photodetector array, with each channel corresponding to an Ising node. At the hardware level, the weight w... k The gradient calculation and its exponent updates are completed in real time on the on-chip digital signal processing unit, without entering the central processing unit, to reduce latency. The gradient calculation period is equal to the mode-locking period, typically 5 to 10 microseconds, meeting the industrial cycle time requirement of thousands of matches per second.

[0102] In terms of performance, this invention underwent comparative experiments on a polyolefin production line: the baseline scheme used an equally weighted average gradient, while the improved scheme used an adaptive weighted gradient. The observed index was reaction time, i.e., the delay from the occurrence of a significant change in light intensity signal to the injection of the suppression signal into the solver. The average reaction time of the baseline scheme was 1.8 microseconds, while that of the improved scheme decreased to 0.9 microseconds, a reduction of approximately 50%. Simultaneously, the average local entropy gradient after refinement decreased from 0.14 to 0.09, indicating a smoother overall matching graph and a reduction of an average of one subsequent causal correction. Regarding energy consumption, due to the more timely suppression intervention, the production line avoided unnecessary high-speed feeding attempts, resulting in an additional 2% reduction in unit energy consumption.

[0103] To further illustrate the adaptability of weight updates, consider this example: During a large material change process, the solver's mode-locking signal experiences three sudden drops. The first drop is due to the gradient... The system triggered suppression; the highest weight was increased from 0.15 to 0.28. A second sudden drop occurred 3 microseconds later, but its magnitude was only 40% of the previous one. The weight decayed to 0.19 due to the reduced rate, effectively preventing over-suppression. The third sudden drop did not exceed the set threshold of 0.5, and the system remained silent, demonstrating that the weight decay mechanism can prevent frequent false triggers of gradient metrics.

[0104] Preferably, when the gradient of the output signal of the Ising model solver exceeds a preset threshold, the neuromorphic computational model reduces the corresponding synaptic weights to generate an inhibition signal, and uses the inhibition signal to modulate the Ising model solver.

[0105] In the plastics order matching process, the Ising model solver is responsible for completing the global one-time matching between orders and production lines in a way that minimizes energy consumption. However, since the coupling matrix changes in real time with the order flow, the light intensity signal output by the solver may exhibit a steep slope. If this is not suppressed in time, it can easily push the system into a high-energy-consumption local minimum. This invention constructs a closed-loop feedback using a neuromorphic computing model: when the gradient of the solver output signal exceeds a threshold, a suppression signal is generated using the rapid plasticity of synaptic weights to reduce the coupling strength in real time, prompting the solver to re-mode-lock in a low-energy-consumption smooth region.

[0106] In principle, the light intensity of the Ising solver forms a sequence I(t) with time t. The gradient index is obtained through continuous sampling. like Where η is the threshold, inhibition is triggered. The neuromorphic computational model is implemented using a spiking neural network structure, enabling synaptic weight updates at the sub-millisecond scale. The inhibition signal magnitude ΔJ ij Calculated based on the partial derivatives of the gradient direction and Hamiltonian with respect to the coupling weights:

[0107]

[0108] α is the inhibition coefficient; σ i ,σ j These represent the spin states of the order node and the production line node, respectively. This formula indicates that if the current pairing causes a rapid increase in energy (positive gradient), the synaptic weights immediately decrease proportionally; no modulation occurs when energy decreases, avoiding excessive suppression. In implementation, the suppression path is directly connected to the solver coupling matrix register via an on-chip bus. Weight updates employ exponential decay, ensuring that hardware-implementable discrete updates map to the theoretically continuous formula. Synaptic states are stored in a static random access memory array and equipped with digital multipliers and adders, ensuring a ΔJ cycle is completed within 500 nanoseconds. ij Write.

[0109] To prevent the suppression signal from excessively weakening the coupling, this invention introduces a recovery threshold ∈. If Then synaptic weights are based on Gradual recovery, where β is the coefficient of recovery. These are the initial coupling weights. This ensures stable matching while avoiding prolonged suppression that could lead to the complete deactivation of the matching pair.

[0110] The effectiveness verification consisted of two parts. In the simulation test, 1000 sets of coupling matrices were randomly generated, each with a size of 128×64. After enabling suppression, the average number of iterations of the solver decreased from 73 to 41, and the convergence time was shortened by about 44%. The final energy value decreased by an average of 6%, indicating that the suppression mechanism effectively guided the system away from high energy consumption.

[0111] Production line example: Within an 8-hour production window, the system processed 240 orders. The threshold η was set to 0.5, and the recovery threshold ∈ was set to 0.2. Statistical results showed that suppression was triggered 27 times, with an average suppression duration of 0.8 microseconds; 4 unplanned line changes were eliminated, saving 3.2% of energy. Simultaneously, the average local entropy gradient of the output matching map decreased from 0.14 to 0.09, and the number of subsequent causal corrections decreased from 5 to 3, further reducing the total matching latency.

[0112] The advantages of the neuromorphic inhibition mechanism are as follows: First, by utilizing the plasticity of synaptic voltage thresholds, coupling weight updates can be completed instantaneously at the hardware level without the intervention of the central processing unit; second, through the mechanism of multiplying the gradient direction by the partial derivative, only the edges that cause energy increases are adjusted, while other edges remain stable; third, a recovery threshold is introduced to avoid sparsity imbalance in the coupling matrix caused by long-term inhibition. In the plasticizing order matching scenario, this closed loop can ensure that the matching engine reconstructs the coupling structure in milliseconds, enabling rapid adaptation to fluctuating operating conditions and improving the overall production scheduling robustness and energy efficiency.

[0113] Preferably, the local entropy gradient is obtained by multiplying the Euclidean distance, the predicted energy consumption, and the switching cost by preset weights and then summing them.

[0114] The local entropy gradient is a scalar indicator used in this invention to evaluate the stability and economy of refined matching. It combines three unrelated but equally crucial influencing factors—geometric fit, production energy consumption, and line switching costs—through a weighted summation to obtain a physically meaningful comprehensive value. This indicator serves both the conflict probability calculation in the causal generative graph model and provides differentiable feedback for entropy modulation bootstrapping, thus forming a core component of the invention's closed loop.

[0115] First, let's explain three fundamental quantities. Euclidean distance D ij Describe the order form vector φ i With production line energy flow vector ψ j The distance after being embedded back into Euclidean space in hyperbolic space is a purely geometric quantity; the predicted energy consumption E ij Based on the production line's historical energy consumption model, the system outputs a predicted unit energy consumption after inputting variables such as the target melt flow index and screw torque distribution; switching cost C ij The data is derived from a material change database, taking into account factors such as cleaning time, residual material loss, and difficulty in adsorbing color masterbatch. Since the three variables have different dimensions and scales, direct use could easily lead to difficulties in converging the weight coefficients during training. Therefore, this invention performs linear normalization on each variable before calculation, mapping it to the [0, 1] interval. The normalized values ​​are represented by the same symbol in the specification.

[0116] The formula for calculating the local entropy gradient is:

[0117] G ij =β1D ij +β2E ij +β3C ij

[0118] β1+β2+β3=1

[0119] Among them G ij Let be the local entropy gradient between order i and production line j; β1, β2, and β3 are the distance weight, energy consumption weight, and switching cost weight, respectively. The three weights satisfy a normalization constraint, such that G... ijThe values ​​are both within a stable range and directly correspond to the proportions of each factor. A dual-objective evolutionary algorithm is used to determine the weights: objective one is to minimize the number of historical line changes, and objective two is to minimize the energy consumption variance. A set of non-dominated solutions is obtained through offline training using multiple days of production logs. Then, based on the management strategy, a trade-off point is selected on the "energy consumption priority—production capacity priority" curve as the formal weight.

[0120] In principle, superimposing three simple linear terms instead of multiplying or combining them nonlinearly has two advantages. First, linear weighting keeps the gradient continuously differentiable throughout the training interval, making it easier for the entropy modulation module to fine-tune the hyperbolic curvature through gradient descent. Second, the linear form is easy to audit transparently in the blockchain ledger, allowing direct verification of each contribution without needing to look at the hidden layer parameters.

[0121] In implementation, local entropy gradient calculation is performed within the same GPU tensor pipeline, synchronized with the light intensity sampling of the Ising model solver. Once an order and production line matching pair is generated, the tensor kernel is triggered, reading the distance tensor, energy consumption prediction tensor, and switching cost tensor. Weighted multiplication and element-wise addition are performed in a streaming manner, and the results are then written to the message bus. The entire process has a latency of no more than 200 microseconds, ensuring completion before the next cycle of causal correction.

[0122] To prevent weights from being swayed by extreme samples during training, this invention introduces adaptive truncation: if an input is greater than 0.95, that input is soft-limited to 0.95, and the corresponding weight is automatically reduced by 5%. For example, dyeing orders with extremely difficult material changes can lead to switching costs approaching 1. Truncation can prevent β3 from reaching its maximum, allowing the system to still consider geometric distance and energy consumption.

[0123] Example: A batch of 64 high-color masterbatch orders is matched across 4 production lines. Let β1 = 0.45, β2 = 0.35, and β3 = 0.20. The data for order 25 and production line B is D. 25,B =0.22, E 25,B =0.48, C 25,B =0.77. According to the formula, G... 25,B =0.45×0.22+0.35×0.48+0.20×0.77≈0.445. Simultaneously calculate another candidate production line C, and obtain G. 25,C =0.39. The lower local entropy gradient led the system to select production line C and send (25,C) to the counterfactual check. Ultimately, the order matching remained stable, and no additional cleaning was required. Through 8 hours of field testing, this batch of orders using the method of this invention reduced cleaning by 2 times compared to the traditional energy-priority strategy, saving 90 kg of resin, advancing the schedule by 35 minutes, and reducing energy consumption by 3.6%.

[0124] From a closed-loop perspective, the local entropy gradient not only serves as a measure of matching stability but also provides a basis for dynamic adjustment of hyperbolic curvature. After 24 hours, if the rate of decrease of the local entropy gradient is found to be less than 2%, the entropy modulation bootstrap module will increase the absolute value of curvature, indirectly increasing the distance difference, thereby compressing the high-gradient region and improving the overall stability of the system.

[0125] The refined matching results, local entropy gradient, and production line physical feedback are input into the causal generation graph model to calculate the conflict probability. When both the production line state change rate and the conflict probability exceed the preset threshold, the Ising model solver is modulated again to generate the correction matching results and output the conflict marker.

[0126] After refining the matching, the system needs to determine whether the matching can be stably executed in a real production environment. To this end, this invention proposes a "causal-pulse dual-path" safety check and correction mechanism: first, the conflict probability is calculated using a causal generation graph model; second, combined with the real-time change rate of the physical feedback from the production line, a secondary modulation of the Ising model solver is triggered, and a correction matching result with conflict markers is output.

[0127] Causal generative graph models are a type of directed acyclic graph that encodes structural equations to characterize the causal relationships between order features, production line status, and external disturbances. Each node in the graph corresponds to a random variable, and the edge weights represent the strength of direct causality. The causal graph of this invention contains 25 nodes, covering key factors such as target melt flow index, screw torque, back pressure, raw rubber viscosity changes, and workshop temperature and humidity. The causal edge weights are learned through a criterion of minimizing structural risk, ensuring both sparsity and reflecting the process chain. For example, high screw torque has a positive causal impact on melt flow index fluctuations in historical data, thus there exists a directed edge with a positive weight in the graph. After learning, counterfactual scenarios can be generated for a given input, i.e., while keeping the values ​​of all variables except one constant, potential disturbances are randomly sampled, and the response of the target variable is calculated.

[0128] The set of matching pairs formed by refining the matching results is denoted as The local entropy gradient matrix is ​​denoted as G. ij Causal generative graph model receives With G ij The conflict probability matrix P is generated by combining real-time production line physical feedback Ω(t) (including torque increase, back pressure slope and cylinder temperature gradient). ij The core formula is as follows:

[0129]

[0130] Where S is the number of counterfactual samples, and ΔM (s) For the torque increment in the s-th counterfactual scenario, For back pressure increment, The indicator function is denoted by a threshold δ, determined by historical outlier quantiles; k1, κ2, and κ3 are normalized weights that sum to 1. This expression means that if the weighted sum of physical increments exceeds the threshold in a counterfactual scenario, the matching pair is considered to be potentially conflicting and included in the probability. The rate of change of production line status is represented by the vector Θ(t), whose Euclidean norm is:

[0131]

[0132] If ||Θ(t)|| corresponds to P ij At the same time, they exceed their respective thresholds η Θ With η P If the current matching scheme is deemed to require immediate correction, the system constructs a suppression increment upon meeting the triggering condition.

[0133]

[0134] Where λ is the conflict suppression coefficient, σ i σ j This represents the current spin state. After suppressing incremental writing to the coupling matrix, the Ising model solver performs fast annealing to obtain the spin correction matching result. Simultaneously generate conflict markers Γ ij If P ij >η P If it is, mark it as 1; otherwise, mark it as 0.

[0135] The key to this closed-loop design lies in using physical feedback to "verify" causal scenarios, rather than simply using entropy gradients or conflict probabilities. This has the advantage of avoiding frequent adjustments during periods of production line inertia, ensuring continuous cycle time; it also prevents the accidental elimination of temporarily high-energy-consuming but tolerable matching pairs by relying solely on local entropy gradients.

[0136] Example: Assume that after refined matching, a matching pair is obtained (order 17, production line C), and the local entropy gradient G 17,C =0.21. The causal generative graph model generates 16 counterfactual scenarios for this matching pair, and calculates the conflict probability P. 17,C =0.35. Simultaneously, the torque increase of production line C was measured in real time. back pressure increase cylinder temperature gradient The vector norm ||Θ(t)|| = 0.08. Let the threshold η Θ =0.05, η P =0.3, both values ​​are triggered. The system calculates the suppression increment. After writing the code into the solver and annealing for 0.4 microseconds, the order was moved to production line B after correction. The new conflict probability decreased to 0.12, and the local entropy gradient decreased to 0.13. Conflict marker Γ 17,C Set it to 1 and send it to the blockchain ledger along with the correction and matching results to provide samples for subsequent comparative learning.

[0137] Performance evaluation shows that within a 24-hour window, the conflict detection and correction process of this invention was triggered 43 times, reducing misadjustments by 18 times compared to the baseline scheme that only uses local entropy threshold triggering. It also avoided two product whitening defects caused by temperature inertia. This demonstrates that the "causal scenario + physical feedback dual-threshold triggering" method has higher discriminative power, and when combined with the Ising model solver's sub-millisecond remode-locking, the production cycle time remains unaffected.

[0138] Preferably, the causal generative graph model uses a diffusion-backward inference method to generate counterfactual order scenarios and calculates the conflict probability based on the number of times the mismatch occurs in the counterfactual order scenarios.

[0139] This invention introduces a causal generative graph model for conflict detection and correction after refining the matching process. The causal generative graph model is a generative graph network that combines structural equations with probabilistic diffusion processes. Each node represents a random variable of the order or production line status, and the weight of each directed edge reveals direct causal dependencies. In the plastics order matching scenario, the graph structure covers 25 key variables, including the order's target melt index, target color, delivery priority, production line screw torque, back pressure, barrel temperature, and workshop humidity. Edge weights are obtained through a criterion of minimizing structural risk, and redundant edges can be automatically removed to ensure graph sparsity.

[0140] Unlike traditional sampling methods based on prior distributions, this invention employs a diffusion-backward inference approach to generate counterfactual order scenarios. Diffusion-backward inference originates from a diffusion probability model, where the forward process progressively adds noise to the observed samples to an isotropic Gaussian distribution, while the backward process gradually denoises back to the original distribution by learning the noise conditional probability density function. In this invention, forward diffusion injects Gaussian noise into the order array and production line state array until the information is completely annihilated. Subsequently, the backward network progressively reconstructs the state sequence under causal graph constraints. Because the reconstruction process explicitly uses a causal structure, each sampling step follows a causal conditional distribution, thus the generated counterfactual scenario maintains both physical consistency and diversity among variables.

[0141] Let the set of matching pairs output by the refined matching be... After selecting a matching pair (i,j), the diffusion backpropagation network generates S sets of counterfactual scenarios, each containing a complete order variable vector and a production line variable vector. For each scenario, the system randomly perturbs the production line variables while keeping the order-side variables constant, and then estimates energy consumption, screw torque crosstalk, and product viscosity changes using a production line simulator. A mismatch is considered to have occurred if the simulation result exceeds a threshold in the following criteria:

[0142]

[0143] in ΔM represents the comprehensive risk quantification index corresponding to order i and production line j in the s-th counterfactual scenario. (s) For torque increment, Δμ (s) For viscosity increment, ΔE (s) For unit energy consumption increment, ρ1, ρ2, and ρ3 are normalized weights, and ζ is the safety threshold. The weights are obtained through bi-objective optimization of the objective function: objective one is to minimize the historical mismatch false alarm rate, and objective two is to maximize the true instance capture rate. The conflict probability can be obtained after evaluating all counterfactual scenarios.

[0144]

[0145] in This is the indicator function. When both the collision probability and the rate of change of the real-time production line status exceed their respective thresholds, the system determines that the matching pair is unstable and requires immediate correction.

[0146] The production line condition change rate vector consists of the screw torque gradient, back pressure gradient, and barrel temperature gradient. Its Euclidean norm... If ||Θ(t)|| exceeds the threshold ∈ Θ And the probability of conflict P ij Exceeding the threshold ∈ P This triggers the inhibition process. The inhibition process reduces the coupling matrix elements through a neuromorphic computational model:

[0147] ΔJ ij =-λP ij σ i σ j

[0148] Where λ is the proportionality coefficient, σ i σ j The Ising spin is used. After the new matrix is ​​written into the solver, a fast annealing process can be completed in just 500 nanoseconds, outputting the correction and matching results. Simultaneously, the matching pairs that have experienced suppression will be marked with conflict markers Γ. ij =1, and together with the local entropy gradient and real-time feedback from the production line, it is written into the blockchain ledger to provide supervised samples for subsequent federated comparative learning.

[0149] The advantages of this invention are reflected in the following aspects: First, interpretability: The causal generation graph provides causal paths at the variable level, and the system can provide diagnostic information such as "increased torque sensitivity leads to increased conflict probability," facilitating engineers' tracing of the source. Second, real-time performance: The diffusion reverse network inference uses 16 steps of inverse diffusion with an inference delay of 1.2 milliseconds, completing the judgment before production line inertia. Third, accuracy: Based on production data from the past 30 days, the false alarm rate using the conflict detection and correction process of this invention is 2.3%, which is 4.1 percentage points lower than the baseline scheme that only uses local entropy thresholds for judgment, and the false alarm rate is reduced by 5.5 percentage points.

[0150] In this example, 60 polypropylene orders are produced in batches across 4 production lines. During the refined matching phase, the system inputs 60 × 4 matching pairs into a causal generation graph model. For each matching pair (order 12, production line B), 20 counterfactual scenarios are generated, 7 of which violate the decision rules. Therefore, P... 12,B =0.35. Real-time monitoring of production line B shows a torque gradient of 0.064 kN / s, a back pressure gradient of 0.032 MPa / s, a temperature gradient of 0.018 degrees Celsius / s, and a vector norm ||Θ(t)|| = 0.075. If the threshold ∈ Θ =0.05, ∈ P =0.3, triggering suppression, suppression increment ΔJ 12,B = -0.13. After the solver underwent rapid annealing, order 12 was reassigned to production line A. No further abnormal signals were triggered during the subsequent 6 hours of production, verifying the effectiveness of the corrective action.

[0151] The morphological vector, energy flow vector, and conflict marker corresponding to the correction matching result are written into the distributed ledger. Each node updates the parameters of the neural network, Ising model solver, neuromorphic computation model, and causal generation graph model through comparative learning and federated aggregation. When the cumulative morphological entropy decrease rate is lower than the preset threshold, the non-Euclidean space curvature, signal gradient threshold, and suppressed signal gain are adjusted, and the parameters are updated synchronously.

[0152] Writing the correction matching results into the distributed ledger and performing federated updates accordingly is the final step in the closed-loop optimization of this invention. The goal is to make the model parameters adaptively drift towards the long-term coupling law of "order-production line" while maintaining the global consistency of geometric metric and suppression threshold with an entropy modulation strategy.

[0153] First, let's explain the content written and the ledger structure. After the correction matching results are generated, the system saves three types of information for each order and production line pair: morphological vector, energy flow vector, and conflict flag. The morphological vector and energy flow vector are derived from the final output of the neural network at the order end and production line end, respectively; the conflict flag takes a value of 0 or 1, used to indicate whether the match is judged as high-risk in causality checking. To ensure that the data is immutable and can be read in parallel, this invention uses a blockchain ledger, with each new batch of matching results written as a block. The block header records the timestamp and the hash of the previous block, and the block body stores triple hashes arranged in key-value format: the key is "order number-production line number", and the value is the vector hash and the conflict flag. This design saves on-chain space and ensures that each node can calculate the hash locally using the original text for consistency verification.

[0154] Whenever a new block is added to the ledger, each participating node (generally corresponding to different production line controllers or scheduling servers) immediately pulls and decodes the new triplet, entering the contrastive learning phase. The goal of contrastive learning is to make the morphology vector and energy flow vector in the embedding space more consistent with the distribution of "positive samples are close together, negative samples are far apart," thereby improving the discriminative power of the subsequent coupling matrix. This invention uses temperature-scaled contrastive loss, and the local loss function is written as:

[0155]

[0156] Where φ i Let ψ be the shape vector of order i. j Let N be the energy flow vector of production line j; set P contains all positive sample pairs with conflict markers of 0; set N... n Negative sample pairs are marked with a conflict symbol of 1; τ is a temperature coefficient used to adjust the sharpness of the distribution. After each node calculates the gradient using its local samples, the gradient weights are aggregated through federated averaging to obtain the global parameter update Δθ. Aggregation formula:

[0157]

[0158] Where K is the number of nodes, n k Let n be the number of samples at node k. tot Let Δθ be the total number of samples. k is the local gradient of node k. This update applies simultaneously to four types of models: vector generator networks, Ising model solver coupling mapping, neuromorphic inhibition rules, and causal generative graph structure coefficients.

[0159] To prevent the model from overfitting to short-term conditions, this invention introduces the cumulative morphological entropy S(t) as a global stability index. Morphological entropy is defined as the Shannon entropy of the morphological vector distribution within period T:

[0160]

[0161] Where C is the number of morphological vector clusters, p c (t) represents the probability of cluster c occurring within period T. The entropy decrease rate is calculated by continuous sliding motion of the system. If r S The absolute value of (t) is below the threshold ∈ S This indicates that the embedded spatial distribution converges too quickly or is locked by local patterns, necessitating a "temperature restart" of the geometric and suppression hyperparameters. The restart strategy consists of three parts: increasing the absolute value of the non-Euclidean space curvature by 5% to increase the distance expansion rate; increasing the output signal gradient threshold η by 1 percentage point to reduce the suppression frequency; and increasing the suppression signal gain α by 1 percentage point to strengthen the penalty for high-risk edges. After the synchronous update, the system re-evaluates r in subsequent cycles. S (t), maintaining an active distribution without excessive oscillation.

[0162] Example: 96 federated rounds per day, S(t) is calculated after each round. For the first 80 rounds r... S (t) averages -0.004, which is below the threshold ∈ S = -0.005 does not trigger a restart; the 81st time a large batch of single high-polymer isobutylene was generated due to the order flow, r S When (t) = -0.007, falling below the threshold, the system automatically adjusted the curvature from -1 to -1.05, increased the suppression threshold from 0.5 to 0.51, and increased the gain α from 0.15 to 0.152. After 15 rounds, the morphological entropy recovered to the gradually decreasing range (average -0.0042), proving that hyperparameter tuning effectively avoided embedding space degradation. Simultaneous monitoring showed that the Ising solver only increased the average number of convergences by 5%, without significantly affecting real-time performance.

[0163] The distributed ledger design also brings two practical benefits. First, traceability: the gradient hash corresponding to any parameter update is permanently stored on the chain, allowing for the identification of erroneous training rounds in post-event audits. Second, decentralized node trust: even if a production line controller goes offline, its failure to upload gradients will not affect aggregation security; the system automatically renormalizes according to sample weights.

[0164] Preferably, the distributed ledger uses the Byzantine fault-tolerant consensus algorithm to record the hash values ​​corresponding to the morphology vector, energy flow vector and collision marker, and adds a timestamp to each record.

[0165] In this invention, the distributed ledger acts as a "trusted memory," storing key information about the correction and matching results for retrieval and verification by each production line node during subsequent comparative learning and federated aggregation. The ledger employs a Byzantine fault-tolerant consensus algorithm to ensure consistency even when up to one-third of the nodes fail or engage in malicious behavior. Addressing the characteristics of large data volumes and strong temporal order in high-speed production scenarios, this invention only writes the hashes of three fields—morphological vector, energy flow vector, and conflict flag—to the chain, and adds a millisecond-level timestamp to each record, thus compressing the on-chain load while satisfying non-repudiation requirements.

[0166] Ledger structure. A set of four pairs M is generated after each round of matching. For each pair (i,j) in the set, the system calculates the morphological vector φ. i With energy flow vector ψ j hash value h φ with h φ The SHA256 algorithm is used; the collision marker value is c∈{0,1}. These three elements are combined to form a record. <h φ ,h ψThe block is defined as follows: `<block>, c, t>`, where `t` is the block timestamp (UTC milliseconds). Records are written to the buffer to be packaged, and then the ordering service instantiated using Byzantine fault-tolerant consensus determines the block order. A block hash is generated and broadcast simultaneously with block completion; other nodes verify the hash and then write it to disk. Because only hashes are recorded, the amount of on-chain data is independent of the vector dimension; the actual vector is stored in a local database, and nodes query the original text and calculate the hash for verification when needed, using the key "order number - production line number".

[0167] Byzantine Fault-Tolerant Consensus: The consensus process consists of three phases: proposal, preparation, and commit. In the proposal phase, the sorting service elects a rotating leader, packages buffer transactions into candidate blocks, and multicasts them. In the preparation phase, each node verifies the transaction format and signature validity, and broadcasts a preparation message. If more than 2f preparation messages are received (f being the number of fault-tolerant nodes), the process proceeds to the commit phase, generating a commit message and updating the local blockchain state. This process does not require a network-wide synchronized clock; consistency is guaranteed solely by digital signatures and vote counts.

[0168] The functions of on-chain data are threefold. First, traceability. Any node can reconstruct historical matching trajectories on the chain based on timestamps, and combine this with local vector hash decryption to pinpoint the source of anomalies. Second, federated learning sampling. When an update round arrives, nodes traverse the block, using hash pairs where c=0 as positive samples and hash pairs where c=1 as negative samples, avoiding data skew. Third, partition security. If a malicious node attempts to forge collision markers to improve its load balancing, the hash of its forged record will not match the local vector, resulting in verification failure and rejection.

[0169] Cumulative morphological entropy. To monitor embedding space activity, morphological entropy is calculated every 24 hours:

[0170]

[0171] Where Q is the number of clusters, p q The probability of cluster occurrence. If the morphological entropy decrease rate... Below the threshold ∈ S The system synchronously increases the absolute value of hyperbolic curvature, the signal gradient threshold, and the suppression signal gain. The parameter modification hash is also written to the chain to ensure that the hyperparameters of all nodes are consistent.

[0172] In an example, a production line node records 12,000 matching hashes in 8 hours, with each record being approximately 128 bytes in size, requiring only 1.5 megabytes of on-chain storage. If the vectors (128-dimensional morphology and 64-dimensional energy flow, each using 32-bit floating-point) were stored directly, 12 megabytes would be needed, reducing the on-chain load by approximately 8 times. During the experiment, three Byzantine nodes were simulated sending erroneous collision markers, which were rejected by ten normal nodes due to hash mismatches with the vectors, but network consistency remained unaffected. Simultaneously, 32 federation rounds were conducted, with 32 parameter hashes written to the chain. After verification by nodes according to timestamps, the hyperparameters were completely consistent, proving the effectiveness of the consensus and traceability functions.

[0173] Preferably, each node performs a federated averaging algorithm to aggregate and update the neural network parameters, Ising model solver parameters, neuromorphic computation model parameters, and causal generative graph model parameters according to the proportion of local samples, and broadcasts the updated parameters to all nodes.

[0174] To enable the plastics order matching system to adapt to different production line groups and order flows over the long term, this invention places the parameter updates of neural networks, Ising model solvers, neuromorphic computing models, and causal generative graph models within a federated learning framework. The core idea of ​​the federated averaging algorithm is that each node completes local gradient calculations while preserving local data from leakage, and then performs weighted aggregation based on the proportion of local samples. The aggregated global parameters are broadcast through a distributed ledger to ensure consistency across all nodes, protecting production secrets and avoiding central bottlenecks.

[0175] In the specific implementation process, such as Figure 2 As shown, whenever the correction matching result is written to the chain, a new block in the ledger triggers a "synchronization signal". Each node first retrieves the morphological vector hash it wrote within the corresponding time window of the new block, calculates positive and negative sample pairs using the local plaintext vector, and obtains the local gradient vector Δθ of the four models by backpropagating the comparative loss according to temperature scaling. k Let the total number of nodes be K, and the k-th node have n samples. k The total number of samples from all nodes is n. tot The federated average performs a weighted summation at the aggregation end where the sorting service is instantiated:

[0176]

[0177] Where Δθ represents the global gradient increment, Δθ k Let n be the local gradient increment at node k. k Let k be the number of local samples at node k. Weighting factor Ensure that production line nodes with large sample sizes contribute more to global updates, but do not leak their actual vector values; if a node temporarily lacks samples, its weight is automatically reduced to avoid noise.

[0178] Example workflow: Assuming an aggregation cycle of 30 minutes and a total of 5 production line nodes, with each node recording 150, 110, 90, 80, and 70 samples respectively within this cycle, the proportions are 0.3, 0.22, 0.18, 0.16, and 0.14. After collecting 5 sets of local gradients, the system weights them according to their proportions to obtain Δθ, which is then split into vector generation network weight increments, Ising coupling map increments, synaptic plasticity increments, and causal graph edge weight increments according to parameter categories. After verifying the block hash is correct, each node loads the increments and performs buffer synchronization.

[0179] To reduce communication overhead, this invention employs intelligent fragmentation during the broadcast phase: four types of parameters are divided into different channels according to their sensitivity. High-sensitivity parameters (such as convolutional kernel weights) are pushed point-to-point within the encrypted channel, while low-sensitivity parameters (such as gating biases) are broadcast in multicast format. The integration time overhead is controlled within 300 milliseconds, and it does not block the next round of data writing.

[0180] Regarding model stability, this invention introduces morphological entropy decay monitoring. The sliding window length is 24 hours. If the morphological entropy decrease rate is less than the threshold of 0.02, the system considers the embedding space to converge too quickly and triggers hyperparameter self-tuning: the absolute value of curvature increases by 0.05, the output signal gradient threshold increases by 0.01, and the suppression signal gain increases by 0.002. The adjustment information takes effect immediately after being broadcast via the ledger, ensuring that the geometric metric and suppression threshold are consistent across all nodes.

[0181] Performance Evaluation: The system was deployed in an actual polyolefin workshop for 7 days, with 48 rounds of federated polymerization per day. Compared to the single-node centralized training baseline, the average number of convergence rounds decreased from 76 to 42 after day 3, reducing convergence time by 45%; total production line energy consumption decreased by 4.3%, and unplanned line changes decreased by 28%. Furthermore, when 2% stochastic gradient noise was injected into the Byzantine nodes, the polymerization end automatically diluted the noise due to its low weight, resulting in a final global parameter trajectory difference of less than 0.5% from the noise-free solution, verifying its robustness.

[0182] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0183] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for matching plastic orders, characterized in that, Includes the following steps: Event encoding is performed on the order parameter stream and production line parameter stream. The neural network generates an order form vector and a production line energy flow vector, forming an initial energy flow potential field. The event encoding measures the adjacent time difference of continuously sampled signals within a sliding window of fixed duration and converts the time difference into a pulse sequence. The neural network is configured with a temporal convolutional layer, a gated linear layer, and a global average pooling layer in sequence. The dimension of the output order form vector is larger than the dimension of the production line energy flow vector. The initial energy flow potential field is generated by the weighted integral of the energy flow vector. In non-Euclidean space, a coupling matrix is ​​constructed based on the distance between the morphological vector and the energy flow vector. This coupling matrix is ​​then input into the Ising model solver to generate the first matching result. When the gradient of the Ising model solver output signal exceeds a preset threshold, the Ising model solver is modulated using an inhibition signal generated by the neuromorphic computing model to generate a refined matching result and calculate the local entropy gradient. Each element of the coupling matrix is ​​generated by substituting the distance between the morphological vector and the energy flow vector in hyperbolic space into an exponential mapping function. The exponential coefficient of the exponential mapping function is adjusted in real time according to the curvature of the hyperbolic space. The local entropy gradient is obtained by multiplying the Euclidean distance, the predicted energy consumption, and the switching cost by preset weights and then summing them. The refined matching results, local entropy gradient, and production line physical feedback are input into the causal generation graph model to calculate the conflict probability. When both the production line state change rate and the conflict probability exceed a preset threshold, the Ising model solver is modulated again to generate a corrective matching result and output a conflict marker. The production line physical feedback includes torque increase, back pressure slope, and cylinder temperature gradient. The causal generation graph model uses a diffusion-backward inference method to generate counterfactual order scenarios and calculates the conflict probability based on the number of mismatches in the counterfactual order scenarios. The production line state change rate vector consists of the screw torque gradient, back pressure gradient, and cylinder temperature gradient. The morphological vector, energy flow vector, and conflict marker corresponding to the correction matching result are written into the distributed ledger. Each node updates the parameters of the neural network, Ising model solver, neuromorphic computation model, and causal generation graph model through comparative learning and federated aggregation. When the cumulative morphological entropy decrease rate is lower than the preset threshold, the non-Euclidean space curvature, signal gradient threshold, and suppressed signal gain are adjusted, and the parameters are updated synchronously.

2. The method according to claim 1, characterized in that, The gradient of the output signal of the Ising model solver is obtained by performing a weighted average operation on continuous light intensity samples, and the weighting coefficients are adaptively updated with the rate of change of light intensity.

3. The method according to claim 1, characterized in that, When the gradient of the output signal of the Ising model solver exceeds a preset threshold, the neuromorphic computation model reduces the corresponding synaptic weights to generate an inhibition signal, and uses the inhibition signal to modulate the Ising model solver.

4. The method according to claim 1, characterized in that, The distributed ledger uses the Byzantine fault-tolerant consensus algorithm to record the hash values ​​corresponding to the morphology vector, energy flow vector and collision marker, and adds a timestamp to each record.

5. The method according to claim 4, characterized in that, Each node performs a federated averaging algorithm to aggregate and update the neural network parameters, Ising model solver parameters, neuromorphic computation model parameters, and causal generative graph model parameters according to the proportion of local samples, and then broadcasts the updated parameters to all nodes.

Images