Industrial optimization method based on AI data fusion decision

Abstract

The invention relates to the technical field of artificial intelligence, discloses an industry optimization method based on an AI data fusion decision, and aims to solve the problems of opaque decision, high man-machine interaction delay and disjunction of an interpretation model and edge hardware in an existing industry AI decision system. According to the method, through multi-source heterogeneous data acquisition and nanosecond time synchronization, AI decision inference and disturbance analysis are executed in parallel by utilizing a field programmable gate array, a decision explanation vector is generated in real time, hardware-level arbitration is performed in combination with confidence evaluation and a manual intervention instruction, and a final control instruction is output. According to the method, interpretable and high-reliability closed-loop control of microsecond-level delay is realized. According to the method, the real-time performance, the credibility and the man-machine cooperation efficiency of industrial automation decision making are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence, and particularly relates to an industry optimization method based on AI data fusion decision. BACKGROUND

[0002] Under the background of deep integration of artificial intelligence and industry optimization, intelligent decision systems based on data fusion have become a core technology path to improve the resource scheduling efficiency in key fields such as manufacturing, energy networks and agriculture. Such systems integrate multi-source heterogeneous data (such as sensor time series signals, enterprise resource planning information), use deep learning models to realize dynamic response and optimization control of complex working conditions, and typical applications include precise temperature control of chemical reaction kettles, short-term load forecasting of power grids and on-demand allocation of irrigation resources. However, although AI models have made significant progress in prediction accuracy, their actual landing rate in industrial sites has been less than 50% for a long time, and the root cause is that the existing architecture has failed to solve the three contradictions of explainability, real-time intervention ability and edge hardware adaptability in the decision-making process.

[0003] Among them, the industry-level AI decision system generally adopts an end-to-end black box model architecture, which makes it impossible for operators to trace the decision logic. For example, when the system instruction "turn off the heater H1" is triggered, the on-site engineer cannot determine whether the decision is due to the instantaneous abnormality of sensor S001 or the historical data drift. According to industry research, more than 78% of operators refuse to execute AI suggestions due to the lack of credible explanations, and instead rely on experience, resulting in an average of 12.7% of energy or material waste. Although some studies have introduced SHAP or LIME post-explanation methods, the generation delay is generally more than 2 seconds, and the explanation module is independent of the decision execution link, which cannot support the millisecond-level intervention requirement.

[0004] Specifically, in the prior art, the current industry AI system regards "decision generation" and "explanation generation" as two separate software tasks, rather than building a deterministic closed loop through hardware integration. The specific manifestations are that the explanation request lags behind the decision completion, the modules rely on non-deterministic network protocols, and the manual intervention lacks a confidence feedback mechanism. Therefore, there is an urgent need for an innovative method that reconstructs from the bottom of the edge computing architecture, embeds a lightweight explanation mechanism into the decision execution link, realizes a real-time closed loop of "decision as explanation, intervention as correction", and thus breaks through the trust and performance bottlenecks of industry AI landing. SUMMARY

[0005] In order to solve the technical problems of the existing industrial artificial intelligence decision system, such as non-transparent decision process, high human-computer interaction delay and disconnection between explanation model and edge hardware, the present application provides an industrial optimization method based on AI data fusion decision, aiming to build a deterministic closed-loop system integrating decision, explanation and intervention through deep integration at the hardware level, thereby improving the reliability and credibility of industrial automatic decision.

[0006] The core technical concept of the present application is to abandon the traditional technical route of serial tasks of decision generation and explanation generation as software level, and instead design a parallel processing mechanism based on system-on-chip hardware architecture. The mechanism solidifies the calculation core of artificial intelligence decision model and local explainable model and human-computer intervention arbitration logic in field programmable gate array logic unit, realizes microsecond-level data exchange between modules by using on-chip bus, ensures the synchronicity of decision generation and explanation generation from the physical layer, seamlessly integrates human intervention into the decision closed loop, and finally outputs an optimized control instruction that is real-time, explainable and dynamically arbitrated.

[0007] According to one aspect of the present application, an industrial optimization method based on AI data fusion decision is provided, which comprises the following steps: Through a multi-source heterogeneous data acquisition interface, real-time acquisition of sensor data stream, equipment state data and artificial intervention instructions in the industrial field is realized; the sensor data stream contains physical quantity measurement values, the equipment state data contains operating parameters, and the artificial intervention instructions contain adjustment parameters input by operators.

[0008] Synchronous time stamping based on hardware high-precision clock is performed on all the acquired data streams, the high-precision clock is synchronized with the master clock source through the precise time protocol, ensuring the consistency of all data in time reference, forming a synchronous data frame carrying nanosecond-level precision time stamp.

[0009] The synchronous data frame is transmitted to a parallel perturbation and inference acceleration module, which is deployed in a field programmable gate array; the module first performs feature extraction on the sensor data stream and equipment state data in the synchronous data frame to form an original industrial data feature vector; then, the module performs two operations in parallel: one is to directly input the original industrial data feature vector into a hardware-deployed artificial intelligence decision model to generate a preliminary artificial intelligence decision instruction; the other is to generate a set of perturbation feature vectors in the neighborhood space of the original industrial data feature vector according to a preset perturbation generation algorithm, and input the set of perturbation feature vectors into multiple hardware replication instances of the artificial intelligence decision model in parallel to obtain a set of corresponding perturbation decision outputs.

[0010] The preliminary artificial intelligence decision instruction, the original industry data feature vector, the group disturbance feature vector and the group disturbance decision output are transmitted to a deterministic explanation vector generation module through a system on chip internal bus; the module is also deployed in the field programmable gate array; the module uses the input data to solve a local linear model in real time through a hardware implemented weighted linear regression algorithm, and coefficients of the local linear model constitute a decision explanation vector; the decision explanation vector quantifies the contribution of each dimension feature in the original industry data feature vector to the generation of the preliminary artificial intelligence decision instruction.

[0011] The preliminary artificial intelligence decision instruction, the decision explanation vector and the human intervention instruction marked with the synchronous timestamp are transmitted to a hybrid decision fusion and arbitration module; the module is also deployed in the field programmable gate array; the module first calculates a confidence score of a current artificial intelligence decision according to statistical characteristics of the synchronous data frame and time sequence stability of the decision explanation vector; then, the module dynamically weights and fuses or selectively covers the preliminary artificial intelligence decision instruction according to a preset arbitration rule set, combines the confidence score and the weight of the human intervention instruction, and generates a final control instruction.

[0012] The final control instruction is output to an executor unit in an industry field; at the same time, the decision explanation vector and the confidence score are transmitted to a real-time visualization unit of a man-machine interaction interface, and are used to show decision basis to an operator.

[0013] As an embodiment of the application, the data acquisition through a multi-source heterogeneous data acquisition interface specifically includes: acquiring speed and torque data of equipment through a controller area network bus interface, acquiring measurement data of a temperature and pressure sensor array through a serial peripheral interface, and acquiring a quantized intervention value from a man-machine interaction interface through a memory-mapped input register. The synchronous timestamp marking based on a hardware high-precision clock specifically captures the arrival time of each data packet in a physical layer interface controller of the data acquisition interface and adds a timestamp.

[0014] As an embodiment of the application, in the parallel disturbance and inference acceleration module, the artificial intelligence decision model is an eight-bit integer quantized convolutional neural network model, and the network structure and weight parameters of the model are compiled and solidified as a logic lookup table and a register resource of the field programmable gate array. The disturbance generation algorithm is a Gaussian kernel function sampling algorithm, which is implemented in hardware through a linear feedback shift register and a digital control oscillator deployed in parallel. The number of the plurality of hardware replication instances is two hundred and fifty-six, so that the inference processes of two hundred and fifty-six disturbance feature vectors are completed within one clock cycle.

[0015] As an embodiment of the present application, in the deterministic explanation vector generation module, the hardware-implemented weighted linear regression algorithm is a least square solver based on a coordinate rotation digital computer algorithm. The solver processes input data through a pipeline structure, and the calculation delay is fixed and irrelevant to data content, thereby ensuring the deterministic delay of explanation vector generation. The decision explanation vector is a floating point number vector with the same dimension as the original industry data feature vector, and the absolute value of each element in the vector represents the importance of the corresponding input feature.

[0016] As an embodiment of the present application, in the hybrid decision fusion and arbitration module, the calculation formula of the confidence score is: the confidence score is equal to a preset coefficient multiplied by the signal-to-noise ratio of the input data, plus another preset coefficient divided by the Euclidean distance between the decision explanation vector and the decision explanation vector at the previous moment. The arbitration rule set specifically includes: when the confidence score is higher than a first preset threshold and there is no manual intervention instruction, the final control instruction is equal to the preliminary artificial intelligence decision instruction; when the confidence score is lower than the first preset threshold but higher than a second preset threshold, the final control instruction is a weighted average of the preliminary artificial intelligence decision instruction and the manual intervention instruction, and the weight is positively correlated with the confidence score; when the confidence score is lower than the second preset threshold, the final control instruction is completely determined by the manual intervention instruction.

[0017] As an embodiment of the present application, the on-chip system internal bus is a high-level extensible interface bus, which provides a data transmission delay of less than 100 nanoseconds between the parallel disturbance and inference acceleration module, the deterministic explanation vector generation module, and the hybrid decision fusion and arbitration module. The real-time visualization unit of the human-computer interaction interface renders the decision explanation vector as a bar chart, and each bar of the bar chart represents an input feature, and the length and color of the bar correspond to the contribution size and sign of the feature.

[0018] In summary, the present application includes at least one of the following beneficial technical effects: (1) By hardware parallel implementation of core computing tasks such as decision, explanation, arbitration in field programmable gate array, the non-deterministic delay introduced by operating system scheduling, network protocol transmission, etc. in traditional software architecture is completely eliminated, and the time delay of decision generation and explanation generation is reduced from seconds to microseconds, realizing real-time synchronous output of decision and explanation.

[0019] (2) The present application constructs a hardware-based deterministic closed-loop system from data input to final instruction output, and the manual intervention instruction directly acts on the arbitration module through a low-delay on-chip interface, ensuring that the response time of human-computer interaction meets the stringent requirements of industrial safety control, effectively avoiding production accidents caused by intervention delay.

[0020] (3): This invention accelerates the complex locally interpretable model algorithm through hardware, enabling it to run efficiently on edge devices with limited computing power. This solves the technical problem that existing interpretable models cannot be deployed in real time in industrial settings, and makes each artificial intelligence decision instruction have an immediate and quantitative decision basis.

[0021] (4): This invention innovatively introduces a confidence quantification mechanism based on data quality and interpretation stability, and integrates it with human intervention at the hardware level to achieve intelligent human-machine collaboration. The system can dynamically adjust the weight of human-machine control according to the reliability of its own decisions, which significantly improves the robustness of the entire industry optimization system and the operator's trust. Attached Figure Description

[0022] Figure 1 This is a logical flowchart of the core method of this invention. Detailed Implementation

[0023] This invention provides an industry optimization method based on AI data fusion decision-making. Its core lies in achieving parallel and deterministic integration of artificial intelligence decision-making, locally interpretable model computation, and human intervention arbitration through a system-on-a-chip (SoC) hardware architecture. Using a field-programmable gate array (FPGA) as the physical carrier, this method reconstructs the traditional software-level serial execution of decision generation, interpretation generation, and human-computer interaction into a hardware-level synchronous processing link, thereby completing a complete closed loop from raw data input to final control command output within a microsecond timescale. The specific embodiments of this invention are described in detail below with reference to the accompanying drawings.

[0024] First, for the industrial optimization method based on AI data fusion decision disclosed in this application, step S1 is first performed: sensor data streams, equipment status data and manual intervention instructions from the industrial site are acquired in real time through a multi-source heterogeneous data acquisition interface.

[0025] Sensor data streams include measurements of physical quantities such as temperature, pressure, flow rate, and vibration.

[0026] Equipment status data includes operating parameters such as motor speed, valve opening, current, and voltage.

[0027] Manual intervention commands include adjustment parameters input by the operator through a human-machine interface, such as setting a target temperature or forcibly shutting down an actuator.

[0028] Specifically, the multi-source heterogeneous data acquisition interface includes a controller area network bus interface, a serial peripheral interface, and a memory-mapped input register. The controller area network bus interface is used to receive device status data from the distributed control unit, and its communication protocol follows the ISO 11898 standard with a baud rate set to 500 kbps. The serial peripheral interface is used to connect to a high-precision analog front-end chip to acquire the raw voltage signals of the temperature and pressure sensor array, with a sampling frequency of 10 kHz.

[0029] The memory-mapped input registers are directly mapped to the graphics processor frame buffer of the human-machine interface to capture the quantized intervention values ​​generated by the operator sliding adjustment bars on the touchscreen. These values ​​are encoded as eight-bit unsigned integers. All interfaces are configured with independent data buffer queues to handle sudden data surges.

[0030] Then, step S2 is executed: synchronous timestamp marking is performed on all acquired data streams based on a hardware high-precision clock. The high-precision clock is driven by the master clock source of the precision time protocol. The timestamp marking operation is completed in the physical layer interface controller of the data acquisition interface. Specifically, whenever a complete data packet arrives at the receive pin of the interface controller, the hardware logic immediately latches the current precision time protocol clock count value and appends it as a 64-bit wide timestamp field to the header of the data packet.

[0031] This process does not rely on CPU interrupts or operating system scheduling; it is entirely controlled by a dedicated hardware state machine, ensuring that the timestamp capture latency is constant and less than 50 ns. After this processing, all data streams from different sources are encapsulated into synchronized data frames carrying nanosecond-level precision timestamps. The frame structure includes a timestamp field, a data source identifier, a payload length, and a checksum. The inter-frame interval is uniformly scheduled by a global clock, with a maximum deviation of no more than one hundred nanoseconds.

[0032] Next, step S3 is executed: the synchronization data frame is transmitted to a parallel perturbation and inference acceleration module, which is deployed on a field-programmable gate array. The module first extracts features from the sensor data stream and equipment status data in the synchronization data frame to form a raw industry data feature vector.

[0033] The feature extraction process includes digital filtering after analog-to-digital conversion, zero-point drift correction, unit normalization, and dimension alignment. For example, the raw temperature sensor readings are filtered by Kalman filtering to eliminate high-frequency noise, then subtracted from the ambient reference temperature and divided by the full-scale range to map to the zero-to-one interval; motor speed data is averaged through a sliding window to eliminate instantaneous fluctuations, and then the ratio with the rated speed is used to obtain dimensionless feature values. The final raw industry data feature vector has 64 dimensions, with each dimension corresponding to a standardized key industry parameter.

[0034] After constructing the original industry data feature vector, the parallel perturbation and inference acceleration module executes two operations in parallel. First, the original industry data feature vector is directly input into a hardware-deployed artificial intelligence decision-making model to generate a preliminary artificial intelligence decision instruction. The artificial intelligence decision-making model is an octet-quantized convolutional neural network model, whose network structure includes three convolutional layers and two fully connected layers, with the activation function implemented using a piecewise linear approximation. The model's weight parameters and network topology are pre-compiled into lookup tables and register-transfer level descriptions, and then fixed in the logic resources of a field-programmable gate array after synthesis, placement, and routing. The inference process is completed within one clock cycle; the clock frequency is 200 MHz, therefore the latency for a single inference is five nanoseconds.

[0035] Secondly, based on a pre-defined perturbation generation algorithm, a set of perturbation feature vectors is generated in the neighborhood space of the original industry data feature vectors. These perturbation feature vectors are then input in parallel into multiple hardware replication instances of the artificial intelligence decision-making model to obtain a set of corresponding perturbation decision outputs. The perturbation generation algorithm is a Gaussian kernel function sampling algorithm, mathematically expressed as: for the original feature vector x, the i-th perturbation vector x_i = x + ε_i, where ε_i is a perturbation vector with a mean of zero and a covariance matrix of σ. 2 The random vector obtained by sampling from the multivariate Gaussian distribution of I, where σ is the preset standard deviation and takes the value of 0.05.

[0036] In the hardware implementation, this sampling process is accomplished collaboratively by 256 parallel-deployed linear feedback shift registers and numerically controlled oscillators. Each linear feedback shift register generates a pseudo-random bitstream, which is converted into Gaussian distributed random numbers by a box-Müller transform circuit, and then superimposed onto each dimension of the original feature vector by a fixed-point multiplier and adder. The resulting 256 perturbed feature vectors are simultaneously fed into 256 decision inference kernels, each of which is a hardware replica of the aforementioned artificial intelligence decision model.

[0037] Since all kernels work in parallel, the 256 perturbation inferences are also completed within one clock cycle, outputting 256 perturbation decision outputs, which constitute the perturbation decision output set.

[0038] Then, step S4 is executed: the initial AI decision command, the original industry data feature vector, the group perturbation feature vector, and the group perturbation decision output are transmitted to a deterministic interpretation vector generation module via the on-chip system bus. This module is also deployed on a field-programmable gate array (FPGA). Using the input data, the module solves a local linear model in real time using a hardware-implemented weighted linear regression algorithm. The coefficients of this local linear model constitute a decision interpretation vector.

[0039] The hardware-implemented weighted linear regression algorithm uses a coordinate-rotation digital computer algorithm to solve the least squares problem. Its objective function is to minimize the weighted squared error:

[0040] Where f(x_i) is the output of the i-th perturbation decision, and π(x, x_i) is the similarity weight, defined as a Gaussian kernel function:

[0041] In hardware, the solution process is broken down into four pipeline stages: The first stage calculates the distance between all perturbation vectors and the original vectors and generates the weight π. In the second stage, a weighted design matrix A and a weighted observation vector b are constructed, where the i-th row of A is √π_i · x_i^T, and the i-th element of b is √π_i · f(x_i). The third stage performs QR decomposition on matrix [A|b], which is implemented using Givens rotation; The fourth stage involves solving the upper triangular equation system Rw = Q^T b through reverse substitution.

[0042] The entire pipeline has twelve stages, with a processing throughput of one interpretation vector every twenty nanoseconds. The computation latency is fixed at 240 ns, independent of the input data content, thus ensuring the determinism of interpretation generation. The final output decision interpretation vector is a 64-dimensional floating-point vector. The absolute magnitude of each element represents the contribution of the corresponding input feature to the initial AI decision instruction, with positive or negative signs indicating a promoting or inhibiting effect.

[0043] The S5 step then proceeds: the initial AI decision instruction, the decision interpretation vector, and the human intervention instruction marked with a synchronization timestamp are transmitted together to a hybrid decision fusion and arbitration module. This module is also deployed on a field-programmable gate array (FPGA). The module first calculates a confidence score for the current AI decision based on the statistical characteristics of the synchronization data frame and the time-series stability of the decision interpretation vector.

[0044] The formula for calculating the confidence score is:

[0045] Where C is the confidence score, α and β are preset coefficients with values ​​of 0.6 and 0.4 respectively, γ is a small constant to prevent division by zero with a value of 0.01; SNR is the signal-to-noise ratio of the input data, which is calculated by the energy spectrum of the short-term Fourier transform of each sensor channel in the synchronous data frame; e_t is the decision interpretation vector at the current time, e_{t-1} is the decision interpretation vector at the previous time, and the Euclidean distance between the two reflects the stability of the interpretation—the smaller the distance, the more stable the decision basis and the higher the confidence.

[0046] Next, the module dynamically weights and fuses or selectively covers the initial AI decision-making instructions based on a preset arbitration rule set, combining the confidence score and the weight of the human intervention instruction. The arbitration rule set specifically includes three scenarios: when the confidence score is higher than the first preset threshold (set to 0.85) and there is no human intervention instruction, the final control instruction is equal to the initial AI decision-making instruction.

[0047] When the confidence score is lower than the first preset threshold but higher than the second preset threshold (set to 0.6), the final control command is the weighted average of the initial AI decision command and the human intervention command. The weight is positively correlated with the confidence score, specifically: Final command = C · AI command + (1 - C) · Human command.

[0048] When the confidence score falls below a second preset threshold, the final control command is determined entirely by human intervention, ignoring the AI ​​decision output. This arbitration logic is implemented in hardware using a finite state machine, with state transitions triggered jointly by the confidence score comparator and the human intervention detector, and the state switching latency is less than 10 ns.

[0049] Finally, step S6 is executed: the final control command is output to the actuator unit in the industrial field; simultaneously, the decision interpretation vector and confidence score are transmitted to the real-time visualization unit of the human-machine interface. The actuator unit includes an electric regulating valve, a frequency converter, and a relay array. After receiving the final control command, it executes the corresponding action, such as adjusting the cooling water flow or starting and stopping the heating element.

[0050] The real-time visualization unit of the human-computer interface renders the decision interpretation vector as a horizontal bar chart. Each bar represents an input feature, and its length and color correspond to the magnitude and sign of that feature's contribution—red indicates positive promotion, and blue indicates negative inhibition. The bar lengths are normalized based on the maximum absolute value. Confidence scores are displayed as a circular progress bar, with green indicating high confidence, yellow indicating moderate confidence, and red indicating low confidence requiring manual intervention. All visualizations are updated at a frequency of 50 Hz to ensure that operators can monitor the decision-making basis and system reliability in real time.

[0051] Based on the above methods, this invention also discloses an industry optimization system based on AI data fusion decision-making. This system includes a multi-source heterogeneous data acquisition and synchronization unit, a hardware processing core based on a field-programmable gate array (FPGA), an actuator interface unit, and a human-machine interface unit.

[0052] The multi-source heterogeneous data acquisition and synchronization unit comprises multiple physical interface controllers, corresponding to the controller area network bus, serial peripheral interface, and universal asynchronous transceiver, respectively. Each physical interface controller integrates a timestamp capture register, which is connected to a global precision time protocol clock controller. The precision time protocol clock controller receives the second pulse signal from the BeiDou satellite timing module and generates a low-jitter system clock through a digital phase-locked loop, which is then distributed to all interface controllers. The timestamp capture register automatically latches the clock count value when a rising edge is detected in the start bit of the data packet, without requiring software intervention.

[0053] The hardware processing core based on a field-programmable gate array (FPGA) integrates a parallel perturbation and inference acceleration module, a deterministic interpretation vector generation module, and a hybrid decision fusion and arbitration module. The parallel perturbation and inference acceleration module comprises a main decision inference kernel and a set of slave decision inference kernel arrays, both constructed using the same hardware-based AI decision model. The main kernel processes the original feature vector, while the slave kernel arrays process the perturbation vector. The module also includes a perturbation vector generator, consisting of 256 independent Gaussian noise generators, each containing a 32-stage linear feedback shift register and a fixed-point box-Muller transform circuit. The hardware implementation of the deterministic interpretation vector generation module is a fully pipelined matrix operation unit that sequentially performs data weighting, matrix construction, QR decomposition, and reverse substitution solution operations. Its processing throughput is strictly matched to the output rate of the perturbation inference module to avoid data accumulation. The hardware implementation of the hybrid decision fusion and arbitration module is a finite state machine with states including "AI-led," "human-machine integration," and "human intervention." The state transition conditions are determined by the combinational logic of a confidence comparator and an intervention detector.

[0054] The actuator interface unit uses the industrial Ethernet protocol to encapsulate the final control commands into real-time data packets, which are then sent to the field actuators via dual-redundant fiber optic links to ensure communication reliability. The human-machine interface unit connects to an industrial touchscreen via a high-speed video interface. After receiving the decision interpretation vector and confidence score, a dedicated graphics rendering engine generates a visual display. The display refresh is synchronized with the system's master clock, eliminating visual latency.

[0055] This embodiment achieves deep hardware-level integration of decision generation, interpretation generation, and human intervention through the aforementioned methods and systems. All critical data is transmitted on the on-chip system's internal bus, a high-level scalable interface bus employing point-to-point connections and a credit flow control mechanism to provide data transmission latency of less than 100 nanoseconds between modules. The end-to-end latency of the entire processing chain, from data acquisition to final command output, does not exceed 300 microseconds, meeting the stringent real-time control requirements of industries such as chemical and power generation.

[0056] Meanwhile, each AI decision comes with an instant-generated explanatory vector and confidence score, enabling operators to understand the decision logic and make precise interventions when necessary, fundamentally solving the black box problem and trust crisis of industrial AI systems.

[0057] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0058] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An industry optimization method based on AI data fusion decision-making, characterized in that, include: The sensor data stream, equipment status data, and manual intervention commands from the industrial site are acquired in real time through multi-source heterogeneous data acquisition interfaces. The acquired data stream is timestamped using a high-precision hardware clock to form a synchronized data frame; The synchronous data frame is transmitted to the parallel perturbation and inference acceleration module deployed on the field programmable gate array, the original industry data feature vector is extracted, and the following is executed in parallel: the original industry data feature vector is input into the hardware-based artificial intelligence decision model to generate preliminary artificial intelligence decision instructions, and a set of perturbation feature vectors is generated and input into multiple hardware replication instances of the artificial intelligence decision model to obtain perturbation decision output. The initial AI decision command, original industry data feature vector, disturbance feature vector, and disturbance decision output are transmitted to the deterministic interpretation vector generation module via the on-chip system internal bus. The decision interpretation vector is generated in real time through a hardware-implemented weighted linear regression algorithm. The initial AI decision command, decision interpretation vector, and human intervention command marked with a synchronization timestamp are transmitted to the hybrid decision fusion and arbitration module. Based on the statistical characteristics of the synchronization data frame and the time series stability of the decision interpretation vector, a confidence score of the current AI decision is calculated. The initial AI decision command is arbitrated based on the confidence score and the weight of the human intervention command to generate the final control command. The final control command is output to the actuator unit, and the decision interpretation vector and confidence score are transmitted to the human-computer interaction interface for visualization.

2. The industry optimization method based on AI data fusion decision-making according to claim 1, characterized in that, The data acquisition through the multi-source heterogeneous data acquisition interface specifically includes: acquiring the speed and torque data of the equipment through the controller local area network bus interface, acquiring the measurement data of the temperature and pressure sensor array through the serial peripheral interface, and acquiring the quantitative intervention value from the human-machine interface through the memory-mapped input register. The synchronization timestamp marking based on a high-precision hardware clock is specifically achieved by capturing the arrival time of each data packet in hardware and adding a timestamp in the physical layer interface controller of the data acquisition interface.

3. The industry optimization method based on AI data fusion decision-making according to claim 1, characterized in that, In the parallel perturbation and inference acceleration module, the artificial intelligence decision model is a convolutional neural network model that has been quantized by 8-bit integers. Its network structure and weight parameters are compiled and solidified into the logical lookup table and register resources of the field programmable gate array. The parallel perturbation and inference acceleration module has a preset perturbation generation algorithm, which is a Gaussian kernel function sampling algorithm. In the hardware, it is implemented through a parallel-deployed linear feedback shift register and a numerically controlled oscillator. The number of hardware replication instances is 256, allowing the inference process of 256 perturbation feature vectors to be completed within one clock cycle.

4. The industry optimization method based on AI data fusion decision-making according to claim 1, characterized in that, In the deterministic interpretation vector generation module, the hardware-implemented weighted linear regression algorithm is a least squares solver based on a coordinate rotation digital computer algorithm. The solver processes input data through a pipeline structure, ensuring a deterministic delay in the generation of interpretation vectors; The decision interpretation vector is a floating-point vector with the same dimension as the original industry data feature vector. The absolute magnitude of each element value in the decision interpretation vector represents the importance of the corresponding input feature.

5. The industry optimization method based on AI data fusion decision-making according to claim 1, characterized in that, The on-chip system's internal bus is a scalable interface bus, which provides a data transmission latency of less than 100 nanoseconds between the parallel perturbation and inference acceleration module, the deterministic interpretation vector generation module, and the hybrid decision fusion and arbitration module. The real-time visualization unit of the human-computer interaction interface renders the decision interpretation vector into a bar chart. Each bar in the bar chart represents an input feature, and its length and color correspond to the magnitude and sign of the feature's contribution.

6. The industry optimization method based on AI data fusion decision-making according to claim 1, characterized in that, The process of constructing the feature vector of the original industrial data includes digital filtering, zero-point drift correction, unit normalization and dimension alignment after analog-to-digital conversion of sensor data streams and equipment status data, ultimately forming a 64-dimensional standardized feature vector.

7. The industry optimization method based on AI data fusion decision-making according to claim 1, characterized in that, The objective function of the weighted linear regression algorithm is to minimize the weighted squared error, where the weights are defined by the Gaussian kernel function, which is calculated based on the Euclidean distance between the feature vectors of the original industry data and each perturbation feature vector.

Images