Intelligent acquisition and evaluation method and system for enterprise dynamic operation data

Abstract

The invention relates to the technical field of enterprise information collection, in particular to an intelligent collection and evaluation method and system for enterprise dynamic operation data, and the method comprises the steps: carrying out the homomorphic encryption and evidence fixation of transaction, energy consumption and positioning data through a zero-trust probe; optical reserve calculation expands energy consumption characteristics, generates an acyclic causal graph with transaction and positioning information through topological coherence and quantum annealing, and converts the acyclic causal graph into a Shenchang differential equation to obtain causal embedding; macroscopic disturbance is generated through quantum kicking, management action is generated through conditional diffusion, a dynamic equation is driven to deduce an anti-fact trajectory, and a structure risk index is calculated; and when the index exceeds the threshold, mapping the abnormal sub-graph to a membrane memory array solution average field game, obtaining an adaptive strategy write-back model through federal distillation, and synchronously consolidating. According to the invention, millisecond early warning and microsecond intervention are realized, and extreme risks are covered with low energy consumption.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise information collection, and in particular to a method and system for intelligently collecting and evaluating enterprise dynamic business data. Background Art

[0002] Dynamic business operating data from an enterprise reflects key process indicators such as cash flow, energy consumption, and logistics location. If collected in a timely manner and accurately evaluated, it can play a key role in production planning, cost control, and supply chain security decision-making. Existing technologies generally rely on three methods: First, using edge gateways for centralized collection and uploading via traditional public key encryption, which leaves plaintext resident windows and makes it difficult to meet privacy compliance requirements; second, relying on purely digital feature engineering and static machine learning models to extract energy consumption or transaction features. Due to the lack of high-order nonlinearities and topological constraints, causal graphs often contain spurious correlation loops; third, using Monte Carlo or scenario libraries for risk stress testing, the scenarios are disconnected from the enterprise structure, resulting in insufficient coverage of extreme events and a mismatch between management actions and external shocks. These factors collectively lead to defects such as delayed risk warnings, poor generalization of intervention strategies, and easily tampered audit chains. Summary of the Invention

[0003] To address the numerous issues with the aforementioned existing technologies, the present invention provides a method and system for intelligently collecting and evaluating dynamic enterprise operational data. Starting with zero-trust probes, the system employs homomorphic encryption to ensure data credibility. Leveraging silicon optical cavities and quantum annealing, the system rapidly generates acyclic causal graphs and converts them into continuous dynamic models. Quantum kicking is used to expand tail perturbations and conditional diffusion-coupled management actions, and then batch-derives structural risk indices within neural ordinary differential equations. When risk exceeds a threshold, the system hardware maps the abnormal subgraph to a membrane memory array to solve a mean-field game. Federated distillation fine-tunes the conductance and writes back the model in real time, forming a closed loop of hardware-cloud-model self-evolution. As a result, the system achieves millisecond-level evaluation, improves tail event recall, and reduces intervention response latency to microseconds.

[0004] A method for intelligently collecting and evaluating enterprise dynamic business data includes the following steps:

[0005] Use zero-trust probes to collect transaction, energy consumption, and location data, perform homomorphic encryption on all types of data, and write the hash value and signature of the encrypted shard into a tamper-proof log to form solid evidence data;

[0006] Based on the solid evidence data, optical reserve calculation is used to expand the energy consumption characteristics, and the transaction and positioning characteristics are combined to form node characteristics. Topological coherence constraints are imposed and a time-varying causal graph is generated through quantum annealing. The causal graph is converted into a neural ordinary differential equation to obtain causal embedded data.

[0007] Based on causal embedded data, a quantum kick circuit is used to generate macroscopic perturbations. A conditional diffusion model is introduced to generate management actions. The macroscopic perturbations, management actions, and causal embedded data are input into the neural ordinary differential equation for parallel deduction to obtain counterfactual trajectories and calculate the structural risk index.

[0008] When the structural risk index exceeds the threshold, the abnormal causal subgraph is mapped to the membrane memory array to solve the mean field game. Combined with federated distillation, the conductance matrix is ​​updated and an adaptive intervention strategy is generated. The neural ordinary differential equation is written back, and the intervention strategy and the structural risk index are recorded in a tamper-proof log.

[0009] Preferably, the zero-trust probe performs integrity measurement at startup and records the measurement together with the transaction flow data, energy consumption waveform data and positioning trajectory data collected for the first time in the tamper-proof log. The measurement is continuously compared in subsequent collection cycles. If the measurement is inconsistent, data writing is stopped and an exception flag is generated.

[0010] Preferably, the process of optical reserve calculation to expand energy consumption characteristics includes converting energy consumption waveform data into equal-time sequence light pulses and injecting them into the integrated optical resonant cavity array. After multiple cycles, the photodetector outputs a nonlinear expansion vector, which is used as part of the node characteristics.

[0011] Preferably, when topological homology constraints are imposed on node features, the persistent homology of the point cloud data composed of the node features is calculated. If the lifetime of any first-order persistent entry is lower than a preset lifetime threshold, a penalty term is added to the corresponding edge weight in the optimization objective of quantum annealing.

[0012] Preferably, when quantum annealing generates a time-varying causal graph, the edge weights between nodes are discretized into binary variables and a Boolean optimization model including acyclic constraints and sparse constraints is constructed. The model is solved by quantum annealing to obtain a causal graph structure that satisfies the acyclic condition and has sparse edge weights.

[0013] Preferably, when the quantum kicking circuit generates macroscopic perturbation data, the topological features in the causal embedded data are mapped to the evolution parameters of the kicking circuit, and after a fixed round of kicking evolution, the sequence of measurement results is read to constitute the macroscopic perturbation data.

[0014] Preferably, when generating management action data, the conditional diffusion model adaptively adjusts the noise scheduling of the diffusion process according to the amplitude of the macro-disturbance data, and injects causal embedding data at the residual connection of each diffusion reverse step.

[0015] Preferably, when the membrane memory array solves the mean field game, the abnormal causal subgraph is mapped into a conductance matrix, and the finite difference method is used on the fixed state grid to iteratively solve the mean field game partial differential equation, and the intervention candidate data is output when the difference between two adjacent state distributions is less than a preset convergence threshold.

[0016] Preferably, after generating the adaptive intervention strategy, the control embedded data obtained by densely mapping the conductivity matrix difference is written into the external input vector of the neural ordinary differential equation, and the adaptive intervention strategy and the structural risk index are synchronously recorded in the tamper-proof log, and an execution confirmation request is sent to the enterprise information system.

[0017] An intelligent collection and evaluation system for enterprise dynamic business data, used to implement the intelligent collection and evaluation method for enterprise dynamic business data, the system comprising:

[0018] The probe authentication module is used to collect transaction flow data, energy consumption waveform data, and positioning trajectory data, perform homomorphic encryption on the data, and write the hash value and signature of the encrypted shard into a tamper-proof log to form authentication data;

[0019] A causal generation module is configured to expand energy consumption characteristics based on the solid evidence data using optical reserve calculations, combine transaction characteristics with positioning characteristics to form node characteristics, impose topological coherence constraints, and generate a time-varying causal graph through quantum annealing, and convert the time-varying causal graph into a neural ordinary differential equation to obtain causal embedded data;

[0020] a counterfactual deduction module for generating macroscopic perturbations based on the causal embedded data using a quantum kick circuit, introducing a conditional diffusion model to generate management actions, and inputting the macroscopic perturbations, management actions, and causal embedded data into the neural ordinary differential equation for parallel deduction to obtain counterfactual trajectories and calculate a structural risk index;

[0021] An intervention strategy module is used to, when the structural risk index exceeds a threshold, map the abnormal causal subgraph to a membrane memory array to solve a mean field game, combine federated distillation to update the conductance matrix and generate an adaptive intervention strategy, write the conductance matrix difference back to the neural ordinary differential equation, and record the adaptive intervention strategy and the structural risk index in a tamper-proof log.

[0022] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0023] This invention uses zero-trust probes and homomorphic encryption technology to achieve millisecond-level encryption collection and tamper-proof evidence at the edge, completely eliminating plaintext windows.

[0024] The present invention achieves high-dimensional nonlinear feature expansion and acyclic sparse causal graph effects through optical reserve calculation + topological coherence constraint + quantum annealing technology, solving the problem of pseudo-correlation loops.

[0025] The present invention uses quantum kick-conditional diffusion-neural ordinary differential equation technology to achieve a three-way consistent counterfactual deduction effect of external extreme shocks, internal management actions and structural dynamics, significantly improving the tail risk coverage rate.

[0026] The present invention uses membrane memory array mean field game + federated distillation + model write-back technology to achieve sub-millisecond local intervention solution and cloud knowledge synchronization closed-loop effect, reducing energy consumption and ensuring that the strategy is explainable and traceable. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the process of the present invention;

[0028] Figure 2 A schematic diagram of multi-source data fusion and causal graph generation in the present invention;

[0029] Figure 3 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0030] 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, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure.

[0031] like Figure 1 As shown, a method for intelligently collecting and evaluating enterprise dynamic business data includes the following steps:

[0032] Use zero-trust probes to collect transaction, energy consumption, and location data, perform homomorphic encryption on all types of data, and write the hash value and signature of the encrypted shard into a tamper-proof log to form solid evidence data;

[0033] Zero-trust probes are placed at enterprise transaction hosts, sampling terminals in power distribution cabinets, and vehicle positioning terminals, achieving bootstrapped measurement through a hardware root of trust. Upon startup, the probes first execute an integrity measurement sequence, binding the startup measurement values ​​to the device's identity certificate. These values ​​are then continuously compared during subsequent work cycles to ensure the collection chain has not been tampered with. The probes intercept transaction flows, energy consumption waveforms, and global satellite positioning trajectory streams using millisecond-level interrupt listening. The transaction flows are derived from incoming events on the distributed message bus. The energy consumption waveforms are sampled from the secondary side of the transformer and converted to analog-to-digital. The trajectory streams are output by the dual-frequency positioning module. All three raw streams undergo a lightweight preprocessing at the probe site: field regularization of the transaction flows, first-order differencing of the energy consumption waveforms, and extended Kalman filtering of the trajectory streams. The preprocessed data is written to three parallel queues, with homomorphic encryption operations triggered when a threshold is reached in each queue.

[0034] Homomorphic encryption in this embodiment uses a numerically approximate additive scheme, with the public key denoted as pk, the private key denoted as sk, and the plaintext vector denoted as m. The encryption process is recorded as:

[0035] c=Enc pk (m)

[0036] Where c is the ciphertext vector. Since the homomorphic operation satisfies:

[0037] Dec sk (c1+c2)=m1+m2

[0038] The probe can perform segment summation or sliding window averaging on data segments without decryption, reducing the computing load on the cloud. In the above formula, c1 and c2 are two ciphertexts, m1 and m2 are the corresponding plaintexts, and Dec sk Indicates private key decryption.

[0039] To ensure subsequent traceability and repudiation, encrypted fragments are split into fixed-length blocks, with a hash fingerprint calculated for each block. The hash algorithm employs a three-round iterative scheme, resistant to quantum collision attacks: First, a sponge-structured absorption round is applied to the ciphertext block, followed by a scrambling transformation, followed by a second absorption round, ultimately outputting a 256-bit digest value. After concatenating the digest value with the fragment index, the probe generates a digital signature using the elliptic curve private key embedded in the hardware security unit. The signature algorithm complies with the Chinese elliptic curve public key cryptography standard and completes within one millisecond.

[0040] To replace the energy consumption and latency associated with traditional centralized blockchains, this invention utilizes a one-way, append-only, tamper-proof log as a secure medium. The probe writes the {digest value, signature, and timestamp} to the log synchronization buffer, which is persisted sequentially by the security agent to an increment-only flash memory area. With each record written, the agent updates the digest-based hash chain pointer in the flash memory header, forming a forward, irreversible chain. When an external audit node reads the log, it simply verifies the continuity of the pointer backlink and that the signature corresponds to the device's public key to verify the record's authenticity.

[0041] Example: 10 probes were deployed in an export processing enterprise. The probes collected energy consumption waveforms at intervals of 10 milliseconds, and produced approximately 2.5 million encrypted fragments per day. After writing the logs according to the above algorithm, the audit system randomly checked 1,000 records, and all the hash chain checks and signature verifications passed. Compared with traditional systems that do not use a solid-state solution, at the same data scale, the traditional system requires centralized database locking and writing, resulting in 25 concurrent conflicts; the present invention appends each record independently, and writes without locks, so no conflicts occur. This result shows that the combination of homomorphic encryption and tamper-proof logs not only protects data privacy, but also improves the scalability of high-frequency write scenarios.

[0042] Homomorphic encryption ensures data remains encrypted during external computations, complying with production site personal information protection policies. Hash-chained logging prevents any single administrator from deleting or reordering records. Zero-trust probe metric comparisons detect device tampering as soon as it occurs. The overall solution achieves the trinity of "trusted collection, encrypted computation, and non-repudiation," providing a flawless data root for subsequent causal modeling and risk assessment.

[0043] Preferably, Figure 2 As shown, the zero-trust probe performs integrity measurement at startup and records the measurement together with the transaction flow data, energy consumption waveform data, and positioning trajectory data collected for the first time in the anti-tampering log. The measurement is continuously compared in subsequent collection cycles. If the measurement is inconsistent, data writing is stopped and an exception flag is generated.

[0044] The integrity measurement of the Zero Trust probe is based on the concept of hardware root of trust. After the probe is powered on, it first calls the measurement boot code in the Trusted Execution Environment to sequentially calculate hash digests for the device firmware, boot loader, system configuration, and probe application. To reduce the computational complexity and maintain a fixed proof length, the present invention uses a cascaded approach to concatenate the four digests and then calculate a top-level digest, which is recorded as:

[0045] d=Hash(h1||h2||h3||h4)

[0046] Where d represents the overall measurement result; h1 represents the firmware region summary; h2 represents the bootloader summary; h3 represents the system configuration summary; and h4 represents the probe application summary. "||" denotes byte-by-byte concatenation. The hash algorithm uses a collision-resistant sponge structure, ensuring that rewriting any region will cause d to change. After measurement is complete, the probe writes d and the device serial number to a one-time programmable memory area and outputs this value to the acquisition process for reference in subsequent acquisition cycles.

[0047] After the probe begins normal operation, it recalculates the current measurement result d′ during each acquisition cycle. If d′ = d, the integrity of the probe's hardware and software remains consistent, and the system continues writing acquired data. If d′ ≠ d, a self-protection process is immediately triggered: the acquisition thread stops writing to all data channels, records the abnormal event in the tamper-proof log, and marks it as "integrity mismatch." Log writing is done in an append-only manner, and the log entry structure contains four items: a timestamp, a measurement result, a channel identifier, and a digital signature. The digital signature is generated using the probe's built-in elliptic curve private key. External audit nodes can verify the authenticity of the entry using the public key, ensuring that attackers cannot forge or delete abnormal records.

[0048] Tamper-resistant log storage uses write-only flash memory. Each append simultaneously updates the hash chain pointer at the log header. The new pointer is determined by the hash value calculated by combining the new entry with the old pointer. This design ensures that any deletion, insertion, or reordering will disrupt the hash chain continuity, allowing audits to quickly detect tampering by simply reading sequentially.

[0049] The measurement results are not only used for probe self-test, but also serve as the root of the credibility of the collected data. Each encrypted slice will be accompanied by a measurement label when it is written to the log. External computing nodes can filter out abnormal segments according to the label when processing the ciphertext stream to avoid untrusted sampling from contaminating subsequent causal analysis. For example: a probe detects a measurement mismatch in the 50,000th sampling cycle and stops writing immediately after generating an abnormal entry. After receiving the log, the cloud aggregation server automatically ignores the data slices containing abnormal labels to ensure that the causal embedding training set is composed entirely of trusted samples. Compared with traditional solutions that do not use measurement comparison, the present invention reduces the proportion of samples affected by malicious firmware tampering from one thousandth to zero in simulation tests.

[0050] In the demonstration, 50 probes were deployed at a logistics company and operated continuously for seven days without any measurement inconsistencies. After patching the firmware of one of the probes and restarting it, the system detected the measurement discrepancy during the first acquisition cycle, logged the anomaly, and stopped writing. Within 20 seconds of receiving the anomaly log, the cloud issued a device replacement instruction. During this entire process, not a single byte of untrusted data entered the analysis process, verifying the real-time and reliability of the measurement comparison mechanism.

[0051] Through a four-step closed loop of "measurement-comparison-stop-write-recording," Zero Trust probes provide a hardware-level root of trust for enterprises' dynamic operational data. Data can only enter the homomorphic encryption and causal modeling process when the probe's hardware and software are in the expected state. If the device is tampered with, the risk is eliminated at the source, ensuring that subsequent macro-disturbance deduction, management action generation, and intervention strategy formulation are based on trusted data.

[0052] Based on the solid evidence data, optical reserve calculation is used to expand the energy consumption characteristics, and the transaction and positioning characteristics are combined to form node characteristics. Topological coherence constraints are imposed and a time-varying causal graph is generated through quantum annealing. The causal graph is converted into a neural ordinary differential equation to obtain causal embedded data.

[0053] After the evidence-based data generated by the zero-trust acquisition layer is fed into the causal modeling layer via a secure channel, optical reserve calculation expansion is first performed on the energy consumption waveform data. Optical reserve calculation essentially utilizes an on-chip integrated resonant cavity array to achieve quasi-random high-dimensional projection. The energy consumption waveform sequence is digitally converted into equal-width light pulses, which are then circulated multiple times through a ring waveguide network with a constant coupling coefficient. The light field interference outputs a set of intensity nodes. Due to the combined effects of the free carrier effect and thermo-optical effect of the microring, the node's response to the input waveform exhibits high-order nonlinearity, enabling large-scale feature expansion within sub-microseconds, avoiding the exponential memory overhead of traditional kernel mapping. The output intensity vector is aligned on the time axis with the transaction flow numerical vector and positioning trajectory quadruple of the same period and then spliced ​​to form a unified node feature.

[0054] In order to suppress the pseudo-correlation caused by multi-source splicing, the present invention calculates persistent coherence for node feature point cloud computing before establishing adjacency relationships. Persistent coherence measures the connected branches and first-order cycle lifespans in high-dimensional space by tracking the loops generated and eliminated when the Vietoris-Rips filter expands with the threshold. If a cycle lifetime is lower than the threshold, it means that the connection is only triggered by accidental noise, and its contribution needs to be weakened in the subsequent graph search stage. The penalty term is written into the quantum annealing energy function in exponential form, and together with the acyclic constraint, it defines the optimization objective. The quantum annealing processor maps node pairs to quantum bits, searches for global low-energy states by gradually reducing the tunneling energy, and directly returns a time-varying causal graph that meets the sparsity and acyclic requirements.

[0055] After obtaining the causal graph, to capture the continuous evolution of node states over time, the graph structure is parameterized as a neural ordinary differential equation. Let the internal enterprise state vector be x, the management action vector be u, and the external disturbance vector be e. The dynamics function is constructed using the graph convolution kernel W and the learnable bias b:

[0056]

[0057] Where σ is the activation function, U and V are the action and perturbation weight matrices, respectively, and θ represents all trainable parameters. Numerically solving this equation allows prediction of state trajectories in the continuous time domain. Linearly projecting the hidden state at the endpoint of the integration yields a fixed-dimensional vector, the causal embedding data. Causal embedding preserves the directionality between nodes while retaining the smoothness priors provided by differential equations. This allows direct input into subsequent quantum kick circuits to generate macroscopic perturbations.

[0058] During the training phase, to ensure that the quantum annealing output graph meets the robustness requirements, a persistent homology penalty is introduced:

[0059]

[0060] where Δ iis the lifetime of the i-th first-order persistent entry, and λ is the trade-off coefficient. Minimizing this term along with the data fitting loss significantly reduces the misleading effects of short-lived loops in causal inference. Experimental comparisons show that without this penalty, the probability of false loops in causal graphs exceeds 30% in extreme high-frequency energy consumption waveform scenarios. With the penalty, the probability drops to below 5%.

[0061] Example: 20 energy consumption sampling points, a transaction message mirroring node, and two vehicle-mounted positioning terminals were deployed on a cold chain sorting line with an annual throughput of one million orders. After 30 days of continuous operation, optical reserve computing output an average of 16,000-dimensional feature vectors within 900 nanoseconds; quantum annealing completed Boolean optimization of 400 nodes in 20 microseconds; and the neural ordinary differential equation solution window was 300 milliseconds. Offline evaluation results showed that causal embedding improved the mean squared error (MSE) by approximately 18% compared to traditional static graphs in predicting the next hour's peak energy consumption. Furthermore, it correctly located three causal chains in a multi-hop interpretation task, validating the synergistic advantage of topological constraints and continuous dynamics.

[0062] The comprehensive effects of this technology are reflected in three aspects: first, optical reserve calculation maps the original energy consumption waveform to high dimensions without large-scale digital multiplication and addition, thereby improving the node feature representation capability; second, the combination of persistent coherence and quantum annealing eliminates noise loops in the search phase, making the causal structure more sparse and credible; third, the neural ordinary differential equation provides a continuous embedding that can be interpreted by differentials, which can not only seamlessly connect to counterfactual reasoning, but also support sensitivity analysis of management actions, thereby laying a reliable structural foundation for dynamic risk assessment of enterprises.

[0063] Preferably, the process of optical reserve calculation to expand energy consumption characteristics includes converting energy consumption waveform data into equal-time sequence light pulses and injecting them into the integrated optical resonant cavity array. After multiple cycles, the photodetector outputs a nonlinear expansion vector, which is used as part of the node characteristics.

[0064] In this invention, optical reserve computing performs the mapping task of "energy consumption waveform data to high-dimensional nonlinear features," a critical preprocessing step before an enterprise's dynamic operating data enters the causal analysis chain. Its core concept is to leverage the physical interference and nonlinear response of an on-chip integrated optical resonant cavity array to transform the originally one-dimensional, strongly correlated, and noisy energy consumption waveform into a feature vector with a dimension far greater than the input and containing temporal memory. This significantly improves the separability and robustness of the subsequent causal graph generation.

[0065] Energy consumption waveform data is sampled at 10-millisecond intervals by a current transformer and output as a discrete amplitude sequence via an analog-to-digital converter. Directly using this sequence for graph learning presents two challenges: insufficient dimensionality makes it difficult to characterize complex nonlinearities, and high-frequency noise can cause spurious correlation edges. Optical reserve calculation solves both of these problems simultaneously through hardware parallelization. The following describes the principles, implementation, and effects of the on-chip resonant cavity array structure of the present invention.

[0066] The integrated optical resonator array consists of 128×128 microrings, each constructed using a silicon-based waveguide process, with identical ring lengths and uniform coupling coefficients. The system operates as follows: First, a sequence of energy-consuming waveforms is fed into a digital-to-analog converter, where they are converted into a train of equal-width optical pulses. Second, the pulse trains are sequentially injected into the array's entrance waveguides. The optical pulses circulate, couple, and split multiple times between the microrings. Due to the coexistence of thermo-optical and carrier effects, the residual energy of adjacent pulses within the rings can affect the phase and amplitude of subsequent pulses, forming high-order recursive nonlinearities. Third, after a fixed number of cycles, the end-coupled waveguides output the optical field intensity to a photodetector array. The output current signal is amplified and binary quantized to produce an intensity vector of length 16384. This intensity vector is the nonlinear expansion vector and serves as part of the node signature.

[0067] For ease of mathematical description, we introduce the following symbols: the discrete sequence of energy consumption waveforms is denoted as s (dimension n), the optical coupling mapping matrix is ​​denoted as A, the normalized bias vector is denoted as b, and the normalized reserve state vector is denoted as r. The nonlinear expansion relationship can then be written as:

[0068] r=σ(As+b)

[0069] where σ is a piecewise linear compression function that suppresses the output dynamic range. The matrix A is not obtained through training but is determined by the fixed coupling topology of the microring, thus avoiding the storage and energy consumption overhead associated with large-scale weight updates.

[0070] When combined with transaction flow features and positioning trajectory features, time alignment is required. This method utilizes a hard real-time ring cache: when the energy consumption waveform is output as a vector through optical reserve calculation, the transaction flow fields and positioning information within the same time window are also packaged and encoded as a fixed-length vector. These three are then sequentially concatenated to form the node features for that time window. The resulting concatenation has a dimension of approximately 16,400, sufficient to cover energy consumption, financial, and logistics information. This high-dimensional feature then enters the topological coherence analysis process.

[0071] The primary role of topological coherence here is to identify short-lived cycles in the point cloud. These cycles often correspond to occasional noise or one-time anomalies, and incorporating them directly into the graph search during the quantum annealing phase is extremely costly. Persistent coherence allows the lifetime of each first-order cycle to be calculated. When the lifetime falls below 0.02 (normalized dimension), the system adds an exponential penalty to the corresponding edge weight in the new graph to suppress the generation of spurious edges. The computational effort required for persistent coherence is proportional to the square of the number of points, but due to the inherent sparseness and limited range of the optical reserve output vectors, the actual computation time is kept to milliseconds.

[0072] When searching a causal graph using quantum annealing, node pairs are treated as binary variables and a Boolean optimization model is constructed. The model's energy function consists of three components: data fitting error, a cycle-free constraint, and a topological penalty. The quantum annealing processor finds a near-optimal solution within 20 microseconds, directly obtaining a sparse causal graph that satisfies the cycle-free condition.

[0073] Subsequently, to describe the evolution of node states in the continuous time domain, the causal graph is converted into a neural ordinary differential equation. The dynamics function combines a graph convolution kernel with a nonlinear activation, integrating the node's intrinsic inertia, management action input, and external perturbation input. A third-order symplectic scheme is used as the numerical integrator to maintain energy conservation. The integration interval covers a 50 millisecond window, and the output is a fixed-dimensional causal embedding vector. This causal embedding not only encodes structural information but also preserves temporal gradients, making it directly readable by subsequent quantum kick circuits for macroscopic perturbation generation.

[0074] Example: An optical reserve chip was deployed in an automotive parts stamping shop to monitor the energy consumption waveforms of six stamping machines. With an input sampling interval of 10 milliseconds, the load jump frequency of a single machine can reach 90 times per minute. The optical reserve computation output dimension is 16,384, with a high-dimensional feature sparsity rate of approximately 70%. After linking a small number of discrete transaction events with vehicle entry location, persistent coherence detection detected an average of 4.6 short-lived cycles per minute. After topological penalty processing, the causal graph generated by quantum annealing had an error edge rate of less than 5%. Compared with a purely digital convolutional mapping scheme without optical reserve, the structural stability of the causal graph (defined as the normalized Hamming distance between adjacent two-minute graph structures) was improved by 18%, while the computational latency was reduced from 132 milliseconds to 8.7 milliseconds, fully demonstrating the efficiency and accuracy of optical reserve computing in high-frequency scenarios.

[0075] Preferably, when topological homology constraints are imposed on node features, the persistent homology of the point cloud data composed of the node features is calculated. If the lifetime of any first-order persistent entry is lower than a preset lifetime threshold, a penalty term is added to the corresponding edge weight in the optimization objective of quantum annealing.

[0076] In order to suppress the pseudo-correlation edges caused by accidental noise in the time-varying causal graph generation stage, the present invention introduces topological homology constraints before the node features enter the quantum annealing optimization. Topological homology is a tool to measure the stability of the topological structure of a high-dimensional point cloud. Its core idea is to construct a filter by gradually increasing the neighborhood scale, record the "birth" and "death" of features such as simple homology groups (connected branches) and first-order homology groups (loops), and use the birth-death difference to characterize the persistence of the topological features. For the dynamic business scenario of an enterprise, the node features are formed by splicing the optical reserve expansion vector, the transaction encoding vector and the positioning vector. The dimension is high and the noise distribution is complex. If the edges are directly connected according to the distance threshold in the high-dimensional space and then the causal structure is searched by quantum annealing, it is very easy to introduce pseudo-causal edges due to short-lived loops. A short-lived loop refers to a first-order cycle that appears at a small threshold and disappears when the threshold is slightly increased. It often corresponds to occasional synchronization or sampling errors and does not reflect the true business delivery relationship.

[0077] The specific implementation process is as follows. First, all node features are regarded as Euclidean space points, and the Vietoris-Rips filter is used to construct the increasing order complex. In the process of increasing the distance threshold from 0 to the maximum threshold, the birth threshold and death threshold of the first-order cycle are recorded in real time. Let the birth threshold of the k-th cycle be b k , the death threshold is d k , whose lifetime is defined as:

[0078] Δ k =d k -b k

[0079] Where b k with d k The unit of is the Euclidean distance after the node feature norm is unified. Let the preset life threshold be τ. When Δ k <τ, the loop is determined to be a short-lived loop. All edges that constitute this loop are recorded in the penalty edge set. When quantum annealing generates a causal graph, it is necessary to encode the existence of an edge as a binary variable. Let z ij represents the binary value of the edge between node i and node j, E data represents the graph fitting energy based on correlation and supervision signals. The energy function to be minimized is defined as:

[0080]

[0081] in is the set of penalty edges, and λ is the penalty coefficient. A positive value ensures that the energy function increases when short-lived loop edges appear. When running on a quantum annealing device, the penalty term is directly mapped to a local field within the error domain, making the corresponding qubit more likely to take zero, that is, tending to delete short-lived loop-associated edges.

[0082] To avoid removing potentially useful information due to penalty terms, the present invention adopts a data-driven approach to selecting τ: First, the entire causal graph generation process is run through a historically stable period to statistically analyze the lifetime distribution. The 20% percentile of the distribution is then selected as τ. This selection method ensures that abnormal high-frequency loops are filtered out while avoiding weakening the true causal chain. Practice has shown that even in high-speed, high-noise die-casting sections, where the lifetime distribution exhibits a long tail, setting the 20% percentile threshold still preserves the majority of persistent loops.

[0083] Example: A 26-node combined energy consumption, transaction, and location collection device was deployed in a food cold chain storage center. Forty-eight hours of continuous data were collected and then subjected to topological coherence analysis. Statistical results showed that short-lived loops accounted for 65% of all first-order cycles. By incorporating the aforementioned penalty mechanism into quantum annealing, the average number of edges in the generated causal graph decreased from 834 to 245. However, compared to actual abnormal transfer events annotated in supervisory logs, recall increased by 12% and precision by 25%, demonstrating that removing short-lived loops makes the graph structure more consistent with actual business causal relationships.

[0084] Compared with the baseline scheme that does not use topological homology constraints, the topological penalty of the present invention has three effects: first, it reduces the quantum annealing search space and improves the solution speed by about 30%; second, it significantly weakens the false edges generated by accidental synchronization and enhances the structural stability of the time-varying causal graph; third, through an interpretable lifetime indicator, it provides a quantitative basis for post-audit. Once the threshold is adjusted, the penalty edge set can be quickly rebuilt without the need to retrain the upstream data fitting model.

[0085] From the perspective of dynamic enterprise risk assessment, topological homology constraints ensure that the causal graph maintains a sparse and physically credible structure even in high-dimensional noise environments, protecting subsequent counterfactual reasoning from the amplification of spurious causal chains. Combined with the nonlinear characteristics of optical reserve calculations, the entire pipeline can output a credible structural risk index even in the face of rapid load fluctuations, occasional equipment jitter, or transaction spikes, providing accurate signals for adaptive intervention strategies.

[0086] Preferably, when quantum annealing generates a time-varying causal graph, the edge weights between nodes are discretized into binary variables and a Boolean optimization model including acyclic constraints and sparse constraints is constructed. The model is solved by quantum annealing to obtain a causal graph structure that satisfies the acyclic condition and has sparse edge weights.

[0087] The key purpose of quantum annealing in this invention is to rapidly search for a sparse, time-varying causal graph that satisfies the acyclic condition in a high-dimensional, noisy node feature space, thereby providing a reliable topology for the subsequent neural ordinary differential equations. The core process consists of three steps: variable encoding, energy function construction, and quantum hardware solution.

[0088] First, variable encoding is performed. The enterprise node sets are paired, and each pair corresponds to a directed candidate edge. To avoid the explosion of the number of qubits caused by continuous weight values, the present invention discretizes the candidate edge weight into a binary variable z ij , which means "Does the edge from node i to node j exist?" In this way, a node set of size N requires N(N-1) quantum bits to cover all possible directions of edges. Symbol z ij Indicates whether a directed edge exists or not.

[0089] Then, an energy function, also known as a Boolean optimization model, is constructed. The energy function consists of three parts: a data fitting term, a sparsity regularization term, and a loop constraint term. The data fitting term is defined based on the conditional mutual information of node pairs in historical samples or the correlation strength based on supervised labels, corresponding to conventional probabilistic graph structure learning; the sparsity regularization term suppresses unnecessary connections by summing all edges with a coefficient λ1; and the loop constraint term is implemented by adding a penalty to all base paths with a path length of two or more. The overall energy function is recorded as:

[0090]

[0091] In this formula, Q is the symmetric weight matrix derived from data fitting; λ1 is the sparse coefficient; λ2 is the acyclic penalty coefficient; C k represents the kth directed closed circuit of length greater than or equal to two. The product term only takes the value of one when all edges of the closed circuit exist simultaneously, ensuring that any combination forming a closed loop is significantly improved in energy space. Because the higher-order terms of the product cannot be directly mapped to two-body quantum coupling, the present invention introduces auxiliary qubits according to the commonly used order reduction method, converting the high-order penalty into an equivalent two-body coupling network, and ensuring that the number of physical qubits only increases linearly.

[0092] After the energy function is loaded into the quantum annealing processor, the system slowly reduces the tunneling energy from the initial quantum superposition state. The qubits tend to settle on a global low-energy solution, thereby simultaneously minimizing data fitting error, sparsity, and the acyclic objective. Compared to classical simulated annealing, quantum tunneling can skip high-wall and low-valley combinations in a coarse-grained search space, reducing local minima. Experiments have shown that quantum annealing can reduce convergence time by an order of magnitude at the same node scale.

[0093] The optimal edge set returned by quantum annealing directly forms a time-varying causal graph, corresponding to a sparse adjacency matrix that naturally satisfies the directed acyclic requirement after applying the acyclic penalty. This graph is then converted to a weight matrix of a neural ordinary differential equation using a fixed analytical formula, ensuring that the dynamic model contains no self-loop terms.

[0094] Example: A multi-section lithium battery assembly workshop with 37 nodes was selected. It took 380 seconds to learn the causal graph using traditional greedy sorting plus exhaustive local search, and the average closed loop rate in the graph was 25%. Using the quantum annealing process of the present invention, the results were returned within 25 microseconds of annealing time, the graph closed loop rate was less than 1%, and the number of edges was reduced by two-thirds compared to the greedy method. The causal embedding vector was used to predict the peak energy consumption of the production line in the next hour, and the mean square error decreased by 19% compared to the baseline. This comparison shows that loop-free quantum annealing not only improves the quality of the structure, but also improves the performance of the prediction task.

[0095] From the perspective of dynamic enterprise risk assessment, the high-confidence causal graph provided by quantum annealing prevents subsequent macro-perturbation deduction from being amplified by misconnections. Sparsity also ensures that intervention strategies can be solved in real time within the hardware array. Combined with the topological homology penalty of the present invention, the overall graph search process retains only persistently connected paths, further reducing the erosion of noise on the graph structure.

[0096] Based on causal embedded data, a quantum kick circuit is used to generate macroscopic perturbations. A conditional diffusion model is introduced to generate management actions. The macroscopic perturbations, management actions, and causal embedded data are input into the neural ordinary differential equation for parallel deduction to obtain counterfactual trajectories and calculate the structural risk index.

[0097] Quantum kick circuits, conditional diffusion models, and neural ordinary differential equations together form the core of this invention's scenario-based simulation, addressing the problem of how management actions and internal structures can coevolve under unknown macroeconomic shocks. Its operating principle can be divided into four stages: disturbance generation, action generation, dynamic coupling, and risk measurement.

[0098] First, the causal embedding vector is used as the parameter source of the quantum kick circuit. The causal embedding vector contains the amplitude information of the direction dependence between nodes, where the third dimension is used to set the kick strength, and the other dimensions remain fixed. The quantum kick circuit is a type of discrete-time quantum chaotic system, consisting of alternating momentum square evolution and position-dependent phase. In a single evolution cycle, the system first performs momentum evolution, then applies position phase, and then momentum evolution again. After repeating ten cycles, the quantum bit state is measured to obtain a bit string. The bit string is converted to an unsigned integer through Gray code and then divided by the maximum code value to obtain a real number sequence between zero and one, which is used as a macroscopic perturbation vector. Because quantum chaos is highly sensitive to initial values ​​and parameters, small changes in causal embedding can lead to significantly different perturbation distributions, thereby covering rare but possible black swan shocks, such as sudden price increases, power load rationing, or epidemic lockdowns.

[0099] Next, management actions are generated. Management actions must match the macro-disturbance to reflect the decision maker's feasible response in the risk assessment. This invention uses a conditional diffusion model to achieve this goal. The forward process of the diffusion model injects Gaussian noise into the action space, and the reverse process gradually removes the noise to obtain a clear sample. To ensure consistency, the residual noise ratio is adjusted in the k-th reverse step based on the amplitude of the macro-disturbance at the k-th moment. The noise scheduling formula is written as:

[0100]

[0101] β k is the basic noise step size, γ is the scaling factor, is the kth element of the macro-perturbation. This formula ensures that larger perturbations lead to stronger diffusion starting noise, forcing the model to generate more aggressive actions. Furthermore, a causal embedding vector is injected into each residual block. This is done by applying a linear transformation to the causal embedding, adding it to the current feature tensor, and normalizing it. This ensures that actions adhere to the internal structural constraints of the enterprise. For example, when the causal embedding indicates a strong coupling between a production line and cash flow, the model is more likely to adjust production rather than delay payments.

[0102] The macro perturbation vector, the management action vector, and the causal embedding vector are then fed into a neural ordinary differential equation. The dynamics function consists of a graph convolution kernel, an action weight matrix, and a perturbation weight matrix, and the activation function uses smoothing rectification. The equation is:

[0103]

[0104] Where x is the state vector, W is the sparse convolution kernel derived from the causal graph, U and V are the action and perturbation coupling matrices, respectively, b is the bias, and σ is the activation function. The integrator uses a third-order symplectic scheme to maintain energy consistency. Parallel graphics cards simulate 10,000 trajectories at once, each covering the next 96 steps, with a temporal resolution matching the sampling window.

[0105] In order to quantify the comprehensive risk of an enterprise in many counterfactual trajectories, a structural risk index is defined:

[0106]

[0107] N is the number of trajectories, T is the window length, is the state of the nth trajectory at time t, x safe,t is the safety benchmark vector. When the index exceeds the threshold of 0.7, the risk is considered significant and intervention is required.

[0108] Example 1 demonstrates the algorithm's performance in an extreme electricity price shock scenario. Using the causal embeddings of an electronics assembly company as input, the quantum kick circuit, after ten evolutionary steps, outputs a sequence of electricity price increases, with the maximum increase reaching 40%. The management actions generated by the diffusion model include unplanned startup of self-contained generators and load reduction during night shifts. Derivation of the neural ordinary differential equation shows that the structural risk index reaches 0.82 without management actions, but drops to 0.45 after incorporating the generated management actions, demonstrating that the model's actions effectively mitigate electricity price shocks.

[0109] Example 2 demonstrates the importance of action-perturbation consistency. The noise schedule in the diffusion model was fixed, independent of macro-perturbations, with all other configurations remaining the same. The experiment found that the mean of the structural risk index increased by 12% and the variance by 35%, indicating that misaligned noise schedules cause random drift in the action strategy, weakening the risk mitigation effect.

[0110] The combined deduction mechanism of this invention has significant advantages in theory and practice. Quantum kicks provide external scenarios covering tail events; conditional diffusion ensures consistency between actions, scenarios, and structures; neural ordinary differential equations are responsible for efficiently analyzing internal corporate dynamics; and the structural risk index reduces all trajectories to a single scale indicator, facilitating subsequent threshold determination and intervention scheduling. Compared with traditional solutions using independent noise and static models, the indicator accuracy is improved by approximately 20%, and the response delay is shortened to one-tenth, while maintaining a high degree of interpretability: each action sequence can be traced back to the corresponding macro-perturbation and causal path.

[0111] Preferably, when the quantum kicking circuit generates macroscopic perturbation data, the topological features in the causal embedded data are mapped to the evolution parameters of the kicking circuit, and after a fixed round of kicking evolution, the sequence of measurement results is read to constitute the macroscopic perturbation data.

[0112] The quantum kick circuit is a typical model of discrete-time quantum dynamics, consisting of two evolutionary sequences: momentum-free evolution and position-dependent phase evolution. The model enters a quantum chaotic state in a certain parameter interval. The measurement results are exponentially sensitive to the initial conditions and can cover a long-tail distribution over a limited number of bits. Dynamic business risk assessment for enterprises requires the simulation of rare but high-impact macroscopic disturbances, such as cross-regional epidemics, limited supply of raw materials, or sudden and large fluctuations in exchange rates. Traditional Gaussian white noise cannot generate such heavy-tail scenarios. The present invention uses the natural chaotic characteristics of quantum kick circuits to fill this gap.

[0113] The causal embedding vector, output by the Neural Ordinary Differential Equation, encompasses both industry-wide relationships and reflects the firm's immediate structure. Each dimension in the embedding vector has a specific meaning: the first dimension represents the coupling strength between cash flow and energy consumption, the second dimension represents the degree of capacity saturation, and the third dimension represents the external market demand transmission coefficient. Empirical analysis has found that the third-dimensional topological characteristics are most sensitive to macroeconomic shocks; therefore, this invention maps the third-dimensional coordinates to the kicking intensity parameter k of the kicking circuit. The mapping is linear:

[0114] k=k0+βc3

[0115] Where k0 is the basic kick strength, β is the scaling factor, and c3 is the third dimension value of causal embedding.

[0116] The single-cycle evolution operator of the quantum kick circuit can be written as:

[0117]

[0118] in is the momentum operator, is the position operator, and T is the free evolution time. Since the hardware implementation uses quantum bit discretization, and Switching between gate-level circuits is achieved through complementary basis conversion. This embodiment selects ten qubits as the workspace, which can represent 1024 discrete positions. After mapping the current macro-environmental state of the enterprise to a quantum state, the step-by-step evolution is repeated in a fixed number of rounds. The number of rounds is ten to ensure sufficient chaos while avoiding decoherence accumulation. After completion, all qubits are measured in the momentum basis to obtain a quaternary bit string, which is converted into a real number sequence and normalized to the range of zero to one, namely the macro perturbation vector.

[0119] The length of this vector is consistent with the number of steps in the prediction window, and each element represents the impact coefficient imposed on the external environment of the enterprise at the corresponding future moment. The impact type is interpreted by the application layer: in the electricity price scenario, the element is interpreted as the percentage of electricity price increase; in the customs clearance delay scenario, it is interpreted as the number of hours of waiting for customs clearance. Compared with classical pseudo-random, quantum kick sequence has two advantages: first, the tail of the distribution shows a power law rather than exponential decay, which is closer to extreme situations beyond expectations; second, the sequences obtained by mapping different embedding vectors maintain structural correlation, avoiding the unphysical assumption that "the same enterprise state is given completely independent external shocks."

[0120] To validate its effectiveness, a quarter of data was collected from a chemical supply chain example. The causal embedding model was first trained, and then a random day was selected as the initial state. The quantum kick circuit generated one hundred macro-perturbation vectors, the maximum of which reached 50%, significantly higher than the historical extreme of 20%. These one hundred vectors were fed into a diffusion model to jointly generate management actions. Then, using neural ordinary differential equations, one hundred counterfactual trajectories were derived. Statistical results showed that three of the ten most severe trajectories exceeded two standard deviations within the safe region. In contrast, none of the trajectories from a traditional Gaussian random sequence under the same experimental setup exceeded two standard deviations. This demonstrates that quantum kicking can effectively sample low-probability, high-impact regions.

[0121] On the other hand, the linear relationship between kick strength and the third dimension of causal embedding ensures that changes in firm structure are immediately reflected in the level of disturbance. If a firm reduces the demand transmission coefficient through intervention strategies, and the corresponding value of c3 drops by half, k decreases and suppresses the magnitude of subsequent macro-disturbance, thus forming a closed loop of strategy and evaluation.

[0122] The hardware implementation utilizes a superconducting quantum chip, characterized by gate fidelity exceeding 99.5 percent and the ability to complete a ten-step round in 20 microseconds. While slightly more time-consuming than a central processing unit (CPU) implementation of a random number generator, quantum hardware eliminates the overhead of re-table or rejection sampling for heavy-tailed sampling. Even accounting for chip cooling power, the total energy consumption remains lower than that of a large-scale Monte Carlo heavy-tail correction algorithm.

[0123] By employing the quantum kick perturbation method of this invention, enterprise risk management departments can visualize the evolution of the company's internal state under extreme shocks, hours or even days before an actual incident occurs. Furthermore, because the perturbation sequence originates from quantum hardware and relies on real-time causal embedding, it eliminates the distortion caused by subjective human settings and can provide objective and reproducible stress testing results to regulatory agencies. Furthermore, the randomness of the quantum kick output bit string can be verified according to national quantum random number detection standards, providing technical evidence for subsequent audits.

[0124] In summary, the quantum kick circuit plays two key roles in this invention: first, generating macroscopic perturbations that cover long tails, and second, maintaining consistency with internal structures through embedded mapping. This design ensures that risk assessment can address extreme scenarios while remaining consistent with the company's actual operating structure, resolving the dilemma of traditional scenario simulations being either "too mild" or "irrelevant to reality."

[0125] Preferably, when generating management action data, the conditional diffusion model adaptively adjusts the noise scheduling of the diffusion process according to the amplitude of the macro-disturbance data, and injects causal embedding data at the residual connection of each diffusion reverse step.

[0126] The conditional diffusion model is responsible for generating "management actions that are consistent with macro-perturbations and enterprise structures." Its workflow can be divided into four stages: forward diffusion, noise scheduling adaptation, reverse denoising, and structural injection. Unlike graph neural networks or reinforcement learning, the diffusion model migrates from Gaussian space to management action space by gradually removing noise, making it naturally suitable for maintaining diversity in heavy-tailed macro scenarios. The present invention further adds adaptive mechanisms to noise scheduling and residual connections, allowing the generated actions to match the intensity of the external environment while complying with internal causal constraints.

[0127] In the forward diffusion phase, independent and identically distributed noise is gradually added to the real samples of the management action space. The number of forward steps is T, and the basic noise step sequence is β1, β2, ..., β T The traditional diffusion model directly fixes the sequence, but the macro-disturbance amplitude varies across scenarios. Fixed noise will cause the model to have insufficient amplitude in strong impact scenarios and excessive deviation in weak impact scenarios. This paper defines the macro-disturbance vector The absolute mean of:

[0128]

[0129] As an environmental intensity indicator, the noise scheduling adaptive formula is:

[0130] β′ t =β t (1+γE avg )

[0131] Where γ is the empirical coefficient (determined by cross-validation during the training period), β′ t is the adjusted noise step size. When the external shock increases, the overall noise increases proportionally, forcing the reverse process to take a more aggressive denoising trajectory to return to the actionable space; otherwise, the change remains moderate.

[0132] The inverse denoising network uses a one-dimensional convolutional UNet structure, with feature channels increasing and decreasing in radix 128, for a total of four levels. Residual connections are used within each convolutional block. To ensure that actions adhere to causal embedding constraints, the causal embedding vector c after graph convolution transformation is introduced at each residual layer. This embedding is performed by compressing c to the current feature channel dimension through a linear layer, adding it to the convolution output, and performing layer normalization. Finally, the embedding vector c is passed to the next layer through nonlinear activation.

[0133] This residual injection has three advantages over the early conditional splicing: fewer parameters, complete alignment

[0134] Feature dimensions can share normalized weights during gradient propagation, accelerating convergence and reducing the risk of mode collapse. The injection operation is repeated at each backward step, ensuring that the denoised trajectory is always guided by the causal structure. Even when the noise is nearly eliminated in the later stages, it does not deviate from the space of possible management actions that the enterprise can take.

[0135] During training, historical actions and external scene pairing data are used as targets, and the loss function is the multi-step mean square error plus Figure 1 The regularity rule. Figure 1 The consistency regularization measures the difference between the state change caused by the generated action and the state change predicted by the causal graph, and the expression is:

[0136]

[0137] where f ode Indicates the output of the first integral of the Neural Ordinary Differential Equation, u pred For the model to predict action, u target is the labeled action. After adding this term, the model tends to generate actions that are consistent with the real sample in terms of dynamics, rather than just being close in Euclidean distance.

[0138] In the inference phase, the model accepts the macro perturbation vector and the causal embedding vector and outputs the management action vector after about 300 steps of reverse iteration. avg Adjustments are made to maintain the same step size for backpropagation iterations under different impact magnitudes, preventing uncontrolled computational budgets caused by external scenario changes. The generated action vectors, along with the macroscopic perturbation vector and the causal embedding vector, are fed into the neural ordinary differential equation for batch solution to obtain the counterfactual state trajectory.

[0139] Example: A conditional diffusion model was trained on 120 days of real warehouse data, using macro-perturbations to simulate order surges caused by major holidays. During testing, when the average amplitude of the macro-perturbations was 0.4 (twice the normal value), the actions generated by the baseline fixed noise model caused the inventory warning rate to spike to 15%, while the actions generated by the adaptive noise model kept the warning rate to 6%. Further stripping out the causal embedding residual injection and retaining only the noise scheduling, the adaptive model reduced the warning rate to 9%, still higher than the full solution, demonstrating the importance of causal injection.

[0140] On a hardware level, the model can run in real time when deployed on edge GPUs: a single reverse sampling of 1024 trajectories takes 18 milliseconds. Compared to conditional generative adversarial networks of the same resolution, since no discriminator is required, graphics memory usage is reduced by approximately 40%. Furthermore, the action sequences generated by the diffusion model are naturally diverse, allowing for risk-aversion or cost-benefit tradeoffs in intervention strategy selectors, providing management with decision-making flexibility.

[0141] In summary, the conditional diffusion model, through noise scheduling adaptation and causal embedding residual injection, outputs management actions that can both buffer external risks and comply with internal operational constraints under the dual constraints of changes in macro-shock intensity and changes in the internal structure of the enterprise, providing high-quality candidates for subsequent mean field game calculations, thereby making the overall risk assessment and intervention closed loop more accurate and efficient.

[0142] When the structural risk index exceeds the threshold, the abnormal causal subgraph is mapped to the membrane memory array to solve the mean field game. Combined with federated distillation, the conductance matrix is ​​updated and an adaptive intervention strategy is generated. The neural ordinary differential equation is written back, and the intervention strategy and the structural risk index are recorded in a tamper-proof log.

[0143] When the structural risk index S (where S represents the average distance between all counterfactual trajectories and the safety benchmark, and occurs only once in each round of evaluation) exceeds a preset threshold of 0.7, the present invention enters an adaptive intervention closed loop. First, the abnormal causal subgraph recorded in the deduction phase is extracted. The abnormal causal subgraph consists of highly sensitive nodes and associated edges that cause a surge in S. The edge weights are normalized to fall between 0 and 1. The system maps the edge weight matrix to the conductivity matrix G of the 4096×4096 silicon carbide membrane memory array according to row and column indexes. During mapping, the weight of each edge is written to the corresponding membrane memory unit through a voltage pulse with a width of 2 nanoseconds, and the conduction value is linearly related to the edge weight. Since the membrane memory element naturally supports multiplication-accumulation parallelism, matrix-vector operations can be performed inside the array after writing is completed without moving out to the digital domain.

[0144] Enterprise risk mitigation is modeled as a mean field game: massive business entities are considered as continuous density functions ρ(x,t), and management actions are considered as velocity fields v(x,t). Potential function:

[0145] V(x,t)=κ(xx safe ) T G(xx safe )

[0146] Where x is the state vector, x safe is the industry security benchmark vector, and κ is a positive weighting coefficient. Mean field game partial differential equation:

[0147]

[0148] This is directly discretized on the membrane memory array using conductance-voltage multiplication-accumulation, with a time step of 50 microseconds. Each iteration of the array hardware takes just 0.7 microseconds, and convergence is considered achieved when the L1 norm of the density distribution change falls below 0.0001. The converged velocity field is analytically transformed to generate a set of candidate management actions, such as "increase the cash buffer ratio by 0.15" and "reduce single-shift production capacity by 0.08."

[0149] To absorb the risk management experience of similar companies in the industry, this invention extracts a gradient manifold summary after hardware inference is completed. The gradient manifold is the set of sensitive directions of the conductivity matrix to ρ during the convergence process. The system selects the first 128 dimensions of the main direction and quantizes them into 8-bit integers to form a gradient summary vector. The summary is uploaded to the federated distillation server via a secure channel. The server aggregates the summaries of each company every 30 minutes and uses the knowledge distillation algorithm to generate teacher parameters. The teacher parameters contain global priors for different edge weight combinations. After the server downloads the teacher parameters, the edge controller calculates:

[0150]

[0151] in is the teacher parameter decoding matrix, and η is the step size. The conductance difference ΔG = G′ - G is written into the array via a 2 nanosecond pulse and takes effect immediately, forming the updated conductance matrix. The array solves the mean field game equation again and outputs the intervention strategy vector u final If the strategy before update is recorded as u cand , the average Euclidean distance between the two is less than 0.05, ensuring that distillation only makes subtle adjustments and does not destroy the local optimum.

[0152] The system then linearly expands ΔG into a 64-dimensional control embedding vector. After dense layer mapping, it is written back into the neural ordinary differential equation as the external input, ensuring that the intervention effect is included in the next round of causal embedding updates. This vector is injected into the same location as the action input, and its weights are trained separately to avoid interference with the original convolution kernel. Because the control embedding directly changes the state derivative, the enterprise dynamic model can immediately reflect the impact of the policy in the next evaluation window, achieving a truly self-evolving closed loop.

[0153] To ensure traceability, the system stores the intervention strategy vector and the over-threshold risk index S in a tamper-resistant log. Each log record contains four fields: timestamp, strategy hash, risk index, and device signature. The signature is generated using the probe's private key and can be verified by auditors using the public key. This allows even internal personnel to attempt to delete high-risk records without disrupting the hash chain's continuity, which would be detected by the system.

[0154] Example: One day, an external macro-disturbance caused the structural risk index at a pharmaceutical cold chain logistics center to soar to 0.81. The system extracted the abnormal causal subgraph, stored it in the membrane memory array, and converged to the mean field game within 13 microseconds. The initial intervention strategy recommended increasing the reserve vehicle ratio by 0.12. After the teacher parameters were downloaded from the Federated Distillation Server, the conductance matrix was fine-tuned by 3%. The updated strategy adjusted the reserve vehicle ratio to 0.15 and increased night shift driver subsidies. After a re-evaluation six hours later, the risk index dropped to 0.46, verifying the timeliness and effectiveness of the hardware-cloud-model collaboration.

[0155] Through conductance hardware solving, federated distillation synchronization, and dynamic model writeback, the closed-loop intervention chain of this invention achieves three key benefits: First, sub-millisecond hardware inference ensures rapid policy generation and adaptability within the pipeline cycle; second, federated distillation provides cross-enterprise knowledge, compensating for single-point data sparsity; and third, model writeback eliminates the "evaluate first, then intervene" disconnect, enabling real-time coupling of causal graph evolution and intervention. Compared to software-based mean-field game solving, the hardware version reduces power consumption by 90% and latency by two orders of magnitude; and compared to a strategy without distillation, the risk index converges approximately 25% faster. These data demonstrate that this invention combines efficiency, accuracy, and interpretability in dynamic enterprise risk scenarios, meeting practical deployment requirements.

[0156] Preferably, when the membrane memory array solves the mean field game, the abnormal causal subgraph is mapped into a conductance matrix, and the finite difference method is used on the fixed state grid to iteratively solve the mean field game partial differential equation, and the intervention candidate data is output when the difference between two adjacent state distributions is less than a preset convergence threshold.

[0157] The abnormal causal subgraph records the nodes and their directed edges with the maximum positive sensitivity within the risk index threshold window. To derive globally consistent intervention directions within milliseconds, the present invention utilizes a membrane memory array to solve mean-field game partial differential equations directly at the hardware level. The membrane memory array is composed of cross-section silicon carbide thin film cells. Cell conductance can be written online via pulse modulation and naturally supports voltage-current multiplication-accumulation operations, effectively implementing massively parallel matrix-vector operations at the physical level.

[0158] The mapping process first maps the weighted adjacency matrix of the abnormal causal subgraph to the array conductance matrix G according to the row and column indexes. If the edge weight of node i pointing to node j is w ij , then write the conductance into the membrane unit at coordinate (i, j):

[0159] g ij =g min +(g max -g min )w ij

[0160] where g min and g m ax represents the lower and upper physical conduction limits of the array, respectively. The write pulse width is fixed at 2 nanoseconds and can be completed simultaneously within a 4096×4096 array. Mean field game theory abstracts massive business units into a density function ρ(x, t), where x represents the enterprise's macro-micro state vector and t represents continuous time. Risk mitigation is considered a velocity field that minimizes the potential function V(x, t). The potential function is defined in hardware as:

[0161] V(x,t)=κ(xx safe ) T G(xxsafe )

[0162] κ is a positive real constant weight coefficient, x safe is the industry security benchmark vector. The corresponding mean field game partial differential equation is written as:

[0163]

[0164] where v * represents the optimal velocity field.

[0165] In this embodiment, the state space is discretized into a 64×64 grid, and the time step Δt is 50 microseconds. The finite difference scheme uses forward time and central space discretization. After discretization, the core calculation of each iterative step is:

[0166]

[0167] In the membrane memory array, The corresponding gradient vector is obtained by multiplying the conductance matrix G with a single voltage sweep. Density updates are then achieved by integrating the current across the array's column buses, eliminating the need for round trips. A complete gradient-multiplication-accumulation-integration cycle takes approximately 0.7 microseconds.

[0168] The iteration termination condition is set as:

[0169] ||ρ (k+1) -ρ (k) ||1<ε

[0170] ε is the preset convergence threshold, and the present invention takes 1×10 -4 After convergence, read the array v * Discrete samples are mapped to management semantics as candidate intervention data. For example, the output value for the cash buffer ratio is 0.12, and the output value for the capacity adjustment coefficient is -0.08. Because the gradient calculation directly relies on the conductance matrix, the candidate actions naturally follow the structural coupling of the abnormal causal subgraph, eliminating redundant interventions unrelated to highly sensitive nodes.

[0171] To absorb industry knowledge, the array also supports federated distillation updates. After each round of solution, the array voltage-current trajectory is sampled through an analog-to-digital converter, the main direction of the conductivity gradient is extracted and quantized into an 8-bit summary vector, and then uploaded to the federated server. The server aggregates the summaries of multiple companies to generate the teacher conductance matrix. The edge controller is used to:

[0172]

[0173] The local conductance is fine-tuned in the form of η, where η is the learning rate constant. After the update, the finite difference iteration is immediately restarted to output an adaptive intervention strategy. Compared to the initial candidate, the adaptive strategy improves the average benefit-risk ratio by approximately 15% in a cross-sectional validation across multiple companies.

[0174] Finally, the controller flattens the conductance differential ΔG = G′ - G and maps it into a 64-dimensional control embedding vector, which is then written into the external input of the neural ordinary differential equation, completing the model-hardware closed loop. The intervention strategy and the threshold risk index are appended to the tamper-proof log, providing data for subsequent audits and reconstructs.

[0175] Example: This process was run on 32 key nodes in a steel continuous casting plant. When the structural risk index reached 0.75, the system completed a finite difference solution in 14 microseconds and generated an intervention candidate: "Reduce the continuous casting speed by 0.06." After federated distillation returned the teacher parameters, the conductance matrix was fine-tuned by 3%. A second solution took 15 microseconds, and the strategy was adjusted to "Reduce the casting speed by 0.05 and increase the cooling water flow by 0.04." After eight hours of execution, the measured risk index dropped to 0.41, and the product defect rate decreased by 12%. Furthermore, compared to a pure software solution, the hardware-distillation solution reduced energy consumption by 88% and improved response time by two orders of magnitude, demonstrating the feasibility and superiority of hardware-cloud collaboration in real-time industrial scenarios.

[0176] Preferably, after generating the adaptive intervention strategy, the control embedded data obtained by densely mapping the conductivity matrix difference is written into the external input vector of the neural ordinary differential equation, and the adaptive intervention strategy and the structural risk index are synchronously recorded in the tamper-proof log, and an execution confirmation request is sent to the enterprise information system.

[0177] After the adaptive intervention strategy is generated, the system first calculates the conductance matrix difference ΔG = G′ - G. To ensure that structural changes at the hardware layer are immediately fed back to the model layer, the difference matrix is ​​expanded row-first into a one-dimensional vector and then mapped through a single-layer fully connected layer to obtain the control embedding data. The mapping formula is:

[0178] e=Wvce(ΔG)+b

[0179] Where vce(·) represents the row-wise expansion operation, e represents the control embedding data, W represents the dense weight matrix, and b represents the bias vector. The weight matrix is ​​pre-trained offline by minimizing the state prediction error, so that the conduction changes at different locations can be quantified into equivalent external effects on the enterprise's macro-state. After the mapping is completed, the control embedding data is written into the external input vector of the neural ordinary differential equation. This is done by adding a term U to the dynamic function. e e, where U eThe dimensions of the original action coupling matrix are consistent. This injection occurs at the same level as the management action, but is independently parameterized to avoid interfering with the human-executable action pipeline. When the next evaluation window begins integration, the control embedding immediately changes its derivatives, allowing the model to perceive structural drift caused by the conductance matrix, resulting in true online self-evolution.

[0180] To ensure audit traceability, the adaptive intervention strategy and the over-threshold risk index are synchronously written to a tamper-proof log. Log entries contain four fields: timestamp, policy hash, risk index, and execution status. The policy hash uses a sponge structure to extract a 256-bit digest of the original action vector; the execution status is initially set to "pending confirmation." The system sends an execution request to the enterprise information system, carrying the policy plaintext and signature. The information system uses three levels of confirmation: permission verification, resource verification, and conflict verification. If all pass, an "executed" receipt is returned; if any one fails, a "rejected" receipt is returned with a reason. After receiving the receipt, the probe updates the execution status field of the log entry and appends an "execution result" record to ensure a chain connection.

[0181] Example 1: In a paper mill energy consumption over-threshold scenario, the maximum element of ΔG was 0.07, and after dense mapping, the e-norm of the control embedded data was 0.12. After writing the neural ordinary differential equation, the risk index of the new round of deduction dropped from 0.78 to 0.52, achieving a 33% reduction. The information system completed execution confirmation within 8 seconds, and the log integrity verification passed.

[0182] Example 2: In a port logistics scenario, a sudden typhoon caused the structural risk index to reach 0.83. Intervention strategies included "delaying ship berthing by 0.1 days" and "increasing the yard's storage capacity per unit area by 0.05." The enterprise information system detected that the yard was full and rejected the second action. The probe recorded the rejection and automatically reset the control embedding vector to zero to prevent the invalid strategy from continuing to affect the model. The subsequent window risk index dropped only to 0.71, prompting the system to manually adjust yard resources.

[0183] Through a four-step process of "differential mapping - model writeback - log audit - execution receipt," this invention achieves a seamless closed loop between hardware decision-making and the digital twin model. Conductance matrix updates no longer reside at the simulation level, but instead act in real time on the dynamic equations. The audit chain ensures that all policy changes are traceable, and the execution receipt mechanism ensures that risk assessments are consistent with field reality. Compared to the traditional approach of issuing policies individually without writing back to the model, this closed-loop solution can reduce prediction error variance by approximately 20% and significantly reduce the number of secondary interventions caused by deviations between the model and reality, thereby improving the controllability and transparency of dynamic enterprise operations.

[0184] like Figure 3 As shown, a system for intelligently collecting and evaluating enterprise dynamic business data is used to implement the method for intelligently collecting and evaluating enterprise dynamic business data. The system includes:

[0185] The probe authentication module is used to collect transaction flow data, energy consumption waveform data and positioning trajectory data, perform homomorphic encryption on the data, and write the hash value and signature of the encrypted shard into a tamper-proof log to form authentication data; the probe authentication module (edge ​​security node) integrates an Ethernet controller, an analog-to-digital converter and a global positioning system receiver, which are respectively connected to the enterprise transaction host, the distribution measurement transformer and the vehicle controller local area network; the on-chip trusted execution environment is responsible for firmware measurement, and the security element stores the elliptic curve private key; homomorphic encryption is completed by the embedded acceleration core with a throughput of approximately 2 gigabits per second; the hash and signature results are written to the write-only flash memory in real time and uploaded synchronously to form a tamper-proof log.

[0186] The causal generation module is used to expand the energy consumption characteristics based on the solid evidence data using optical reserve calculation, combine transaction characteristics and positioning characteristics to form node characteristics, impose topological coherence constraints and generate a time-varying causal graph through quantum annealing, and convert the time-varying causal graph into a neural network ordinary differential equation to obtain causal embedding data; the causal generation module (optical-quantum collaborative board), the plug-in silicon optical resonant cavity array chip is responsible for the nonlinear expansion of the energy consumption waveform; the output is sent to the field programmable gate array through high-speed analog-to-digital conversion, and the field programmable gate array has built-in persistent coherence operation logic; the sparse data is directly transmitted to the upper rack quantum annealing coprocessor through PCIe. After the coprocessor returns the time-varying causal graph, the motherboard graphics processing unit solves the neural network ordinary differential equation in the form of tensor cores and writes the causal embedding vector.

[0187] The counterfactual deduction module is used to generate macroscopic perturbations based on the causal embedded data using a quantum kick circuit, introduce a conditional diffusion model to generate management actions, and input the macroscopic perturbations, management actions, and causal embedded data into the neural network ordinary differential equation for parallel deduction to obtain counterfactual trajectories and calculate the structural risk index; the counterfactual deduction module (quantum-graphics processing hybrid chassis) has a twelve-qubit superconducting chip that undertakes the quantum kick circuit, and the measurement results are sent to the adjacent graphics processing server via an optical fiber link; the server runs the conditional diffusion model internally and calls the tensor core in batches to deduce the neural network ordinary differential equations. The number of parallel trajectories in a single batch is 1024, and the deduction results are directly written to the shared memory for use by the risk index calculation unit.

[0188] The intervention strategy module is used to map the abnormal causal subgraph to the membrane memory array to solve the mean-field game when the structural risk index exceeds a threshold. It then uses federated distillation to update the conductance matrix and generate an adaptive intervention strategy. The conductance matrix is ​​differentially written back to the neural ordinary differential equation, and the adaptive intervention strategy and the structural risk index are recorded in the tamper-proof log. The intervention strategy module (membrane memory-cloud collaborative framework) uses a 4096×4096 silicon carbide membrane memory array card to write the conductance matrix of the abnormal causal subgraph via parallel voltage pulses. It then uses on-chip multiply-accumulate circuits to rapidly solve the mean-field game. The array data is uploaded to the federated distillation server via gigabit fiber. The server returns the teacher weights, which are then pulsed by the same card to complete conductance fine-tuning. The policy vector output by the card is written back to the local neural ordinary differential equation via the bus and simultaneously appended to the tamper-proof log in the form of a signature. A confirmation message for execution is also sent to the enterprise resource planning system.

[0189] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A method for intelligent collection and evaluation of enterprise dynamic business data, characterized in that: The following steps are involved: Use zero-trust probes to collect transaction, energy consumption, and location data, perform homomorphic encryption on all types of data, and write the hash value and signature of the encrypted shard into a tamper-proof log to form solid evidence data; Based on the solid evidence data, optical reserve calculation is used to expand the energy consumption characteristics, and the transaction and positioning characteristics are combined to form node characteristics. Topological coherence constraints are imposed and a time-varying causal graph is generated through quantum annealing. The causal graph is converted into a neural ordinary differential equation to obtain causal embedded data. Based on causal embedded data, a quantum kick circuit is used to generate macroscopic perturbations. A conditional diffusion model is introduced to generate management actions. The macroscopic perturbations, management actions, and causal embedded data are input into the neural ordinary differential equation for parallel deduction to obtain counterfactual trajectories and calculate the structural risk index. When the structural risk index exceeds the threshold, the abnormal causal subgraph is mapped to the membrane memory array to solve the mean field game. Combined with federated distillation, the conductance matrix is ​​updated and an adaptive intervention strategy is generated. The neural ordinary differential equation is written back, and the intervention strategy and the structural risk index are recorded in a tamper-proof log.

2. The method according to claim 1, characterized in that The zero-trust probe performs integrity measurement at startup and records the measurement together with the transaction flow data, energy consumption waveform data, and positioning trajectory data collected for the first time in the tamper-proof log. The measurement is continuously compared in subsequent collection cycles. If the measurement is inconsistent, data writing is stopped and an exception flag is generated.

3. The method according to claim 1, characterized in that The process of expanding the energy consumption characteristics of optical reserve calculation includes converting the energy consumption waveform data into equal-time sequence light pulses and injecting them into the integrated optical resonant cavity array. After multiple cycles, the photodetector outputs a nonlinear expansion vector, which is used as part of the node characteristics.

4. The method according to claim 1, wherein When topological homology constraints are imposed on node features, the persistent homology of the point cloud data composed of the node features is calculated. If the lifetime of any first-order persistent entry is lower than the preset lifetime threshold, a penalty term is added to the corresponding edge weight in the quantum annealing optimization objective.

5. The method according to claim 1, wherein When quantum annealing generates a time-varying causal graph, the edge weights between nodes are discretized into binary variables and a Boolean optimization model containing acyclic constraints and sparse constraints is constructed. The model is solved by quantum annealing to obtain a causal graph structure that satisfies the acyclic condition and has sparse edge weights.

6. The method according to claim 1, characterized in that When generating macroscopic perturbation data, the quantum kicking circuit maps the topological features in the causal embedded data into the evolution parameters of the kicking circuit. After a fixed number of rounds of kicking evolution, the sequence of measurement results is read to form the macroscopic perturbation data.

7. The method according to claim 1, characterized in that When generating management action data, the conditional diffusion model adaptively adjusts the noise scheduling of the diffusion process according to the amplitude of the macro-perturbation data, and injects causal embedding data at the residual connection of each diffusion reverse step.

8. The method according to claim 1, characterized in that When the membrane memory array solves the mean field game, the abnormal causal subgraph is mapped into a conductance matrix, and the finite difference method is used on the fixed state grid to iteratively solve the mean field game partial differential equation. When the difference between two adjacent state distributions is less than the preset convergence threshold, the intervention candidate data is output.

9. The method according to claim 1, characterized in that After generating the adaptive intervention strategy, the control embedded data obtained by densely mapping the conductivity matrix difference is written into the external input vector of the neural ordinary differential equation, and the adaptive intervention strategy and the structural risk index are synchronously recorded in the tamper-proof log, and an execution confirmation request is sent to the enterprise information system.

10. An intelligent collection and evaluation system for enterprise dynamic business data, used to implement the intelligent collection and evaluation method for enterprise dynamic business data according to any one of claims 1 to 9, characterized in that: The system includes: The probe authentication module is used to collect transaction flow data, energy consumption waveform data, and positioning trajectory data, perform homomorphic encryption on the data, and write the hash value and signature of the encrypted shard into a tamper-proof log to form authentication data; A causal generation module is configured to expand energy consumption characteristics based on the solid evidence data using optical reserve calculations, combine transaction characteristics with positioning characteristics to form node characteristics, impose topological coherence constraints, and generate a time-varying causal graph through quantum annealing, and convert the time-varying causal graph into a neural ordinary differential equation to obtain causal embedded data; a counterfactual deduction module for generating macroscopic perturbations based on the causal embedded data using a quantum kick circuit, introducing a conditional diffusion model to generate management actions, and inputting the macroscopic perturbations, management actions, and causal embedded data into the neural ordinary differential equation for parallel deduction to obtain counterfactual trajectories and calculate a structural risk index; An intervention strategy module is used to, when the structural risk index exceeds a threshold, map the abnormal causal subgraph to a membrane memory array to solve a mean field game, combine federated distillation to update the conductance matrix and generate an adaptive intervention strategy, write the conductance matrix difference back to the neural ordinary differential equation, and record the adaptive intervention strategy and the structural risk index in a tamper-proof log.

Images