Heterogeneous computing architecture-oriented adaptive multi-modal building equipment data processing method

Abstract

The invention relates to the technical field of building automation control and data processing, in particular to a heterogeneous computing architecture-oriented adaptive multi-modal building equipment data processing method. The method specifically comprises the following steps: S1, multi-source data classification and modal identification; s2, selecting a self-adaptive preprocessing strategy; s3, parallel filtering and feature extraction; s4, executing a multi-task cooperative control algorithm; s5, data fusion and output instruction generation; and S6, adaptive learning and parameter optimization are carried out. According to the algorithm, processing strategies can be adaptively selected according to data types and modalities, universal processor and special accelerator resources in a heterogeneous computing architecture are reasonably scheduled, and multi-task parallel processing is achieved; and meanwhile, through a dynamic parameter optimization mechanism, the data processing precision and the system control performance are continuously improved, and the real-time control requirement in a complex building scene is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of building automation control and data processing, in particular to a self-adaptive multi-modal building equipment data processing method for heterogeneous computing architecture. The algorithm is specially designed for the processing needs of multi-source and heterogeneous data in building automation systems, can fully adapt to the hardware characteristics of heterogeneous computing architecture, significantly improve the processing efficiency and control accuracy of building equipment data, and can be widely applied to various complex building automation scenes such as large intelligent buildings, data centers, commercial complexes, and industrial plants. BACKGROUND

[0002] With the continuous improvement of building intelligence level, the types of devices connected by building automation systems (BAS) are increasingly diverse, covering air conditioning and ventilation systems, water supply and drainage systems, power supply and distribution systems, lighting systems, security systems, and other subsystems. The data generated by these systems exhibits significant multi-modal and heterogeneous characteristics. These data include analog signals such as temperature, humidity, and pressure, as well as digital signals such as device switch states and fault alarms. In addition, there are bus-type data transmitted through BACnet, Modbus, LonWorks, and other bus protocols. The data scale and complexity grow exponentially.

[0003] The existing data processing algorithms in building automation systems have the following shortcomings: first, the processing flow is fixed, using a "one-size-fits-all" processing mode, lacking adaptive recognition ability for different modal data. For example, the same filtering and preprocessing strategy is used for periodic temperature data and sudden fault alarm data, resulting in low processing accuracy. Second, the preprocessing strategy is single, and the processing method is not dynamically adjusted according to the data type. For example, linearization calibration is not performed on analog signals with noise interference, and effective de-bouncing processing is not performed on digital signals with frequent jitter, affecting the reliability of subsequent data processing. Third, the hardware advantages of heterogeneous computing architecture are not fully utilized. Traditional algorithms run on general-purpose processors, without reasonable scheduling of special hardware resources such as signal filter accelerators and PID operation accelerators, making it impossible to achieve multi-task parallel processing, resulting in high data processing delay and difficulty in meeting real-time control requirements. Fourth, there is a lack of dynamic parameter optimization mechanism. Filter parameters and PID control parameters are mostly manually preset and cannot be adjusted adaptively according to system operating conditions such as device load changes and external environment fluctuations. When the system has steady-state errors or overshoots, the parameters cannot be corrected in time, affecting the control effect.

[0004] To solve the above problems, there is an urgent need in the field for an algorithm that can adapt to heterogeneous computing architecture and has multi-modal data adaptive processing capability to achieve efficient and accurate processing of building equipment data and improve the overall performance of building automation systems. SUMMARY

[0005] The core purpose of the present application is to overcome the defects of the existing building automatic control system data processing algorithm, and provide an adaptive multi-modal building equipment data processing algorithm for heterogeneous computing architecture. The algorithm can adaptively select processing strategies according to data types and modalities, reasonably schedule general processors and special accelerator resources in the heterogeneous computing architecture, and realize multi-task parallel processing. At the same time, through the dynamic parameter optimization mechanism, the data processing accuracy and system control performance are continuously improved to meet the real-time control requirements in complex building scenes.

[0006] Specifically, the technical solutions include the following: An adaptive multi-modal building equipment data processing method for heterogeneous computing architecture includes the following steps: S1: multi-source data classification and modality recognition; S2: adaptive preprocessing strategy selection; S3: parallel filtering and feature extraction; S4: multi-task collaborative control algorithm execution; S5: data fusion and output instruction generation.

[0007] Preferably, the S1 specifically includes: According to the source and transmission characteristics of the data, it is divided into three categories: analog quantity, digital quantity and bus type data; On the basis of classification, modalities of each type of data are identified: for analog quantity data, the period characteristics, mutation frequency and variance value of the signal are calculated to identify it as a periodic signal, a mutation signal and a steady-state signal; For digital quantity data, the duration, frequency of change and coding rules of the signal are analyzed to identify it as a switch quantity, pulse count and state quantity; For bus type data, the header identification, protocol field and check bit of the data frame are analyzed to identify the protocol type and data frame structure, and the valid data in the frame is extracted; The identification results are stored in the storage module simultaneously to provide a basis for subsequent processing strategy selection.

[0008] Preferably, the S2 specifically includes: According to the classification and modality recognition results of the S1, different types of data are dynamically matched with preprocessing strategies, and the preprocessing process is completed by the general processor and the storage module.

[0009] Preferably, the preprocessing strategy is as follows: 1) Analog quantity data preprocessing: According to the characteristics of analog quantity data being susceptible to noise interference and having nonlinear errors, a combined strategy of "linearization calibration- outlier rejection- dimensionless normalization" is adopted; Linearization calibration converts the collected voltage / current signal into actual physical quantity by querying the sensor calibration curve, and corrects the non-linear error of the sensor; the abnormal value elimination adopts 3 σ criteria, calculates the mean and standard deviation of the data, and determines the data exceeding the range of [mean-3 σ , mean+3 σ ] as abnormal value and eliminates it, avoiding the influence of extreme data on subsequent processing; dimensionless normalization converts different physical quantities to the interval [0, 1], the formula is: normalized value=(original value-minimum value) / (maximum value-minimum value), which provides a unified dimension for subsequent feature extraction and data fusion; 2) Digital data preprocessing: in view of the problems of jitter and non-uniform state coding of digital data, the "de-bouncing processing-state coding" strategy is adopted; De-bouncing processing sets a time window, if the signal state continuously remains the same within the window, it is confirmed as a valid state, otherwise it is considered as jitter and ignored; state coding converts the discrete states of different devices into uniform binary coding, improving the standardization of data processing; 3) Bus type data preprocessing: in view of the characteristics of protocol difference and data transmission error of bus type data, the "protocol analysis-data validity check" strategy is adopted; Protocol analysis calls the corresponding protocol analysis module according to the protocol type identified by S1 to extract the key information in the data frame; data validity check judges whether the data appears packet loss or error code in the transmission process by checking the check bit of the frame tail, if the check fails, it requests retransmission to ensure the integrity of the data.

[0010] Preferably, the S3 specifically comprises: The preprocessed data of various types by S2 is dispatched to the signal filtering accelerator of the heterogeneous computing architecture by the general-purpose processor to realize parallel filtering and feature extraction; the signal filtering accelerator adaptively selects the filtering method according to the analog quantity data mode identified by S1.

[0011] Preferably, the filtering method comprises: 1) Periodic signal filtering: Kalman filtering algorithm is adopted to dynamically estimate the true value of the signal by establishing the state equation and observation equation of the signal; 2) Abrupt signal filtering: adaptive threshold filtering algorithm is adopted to dynamically adjust the threshold according to the historical fluctuation of the signal; 3) Steady-state signal filtering: sliding average filtering algorithm is adopted, a sliding window with length M is set, the average value of the sampling data in the window is taken as the filtering result at the current time, the formula is: Y( k )=(X( k )+X( k -1)+…+X(k M+1)) / M; After the filtering is completed, the signal filtering accelerator synchronously extracts key features of various data; the extracted feature data is stored in a storage module to provide support for subsequent control operation and data fusion.

[0012] Preferably, the S4 specifically comprises: The general processor assigns a control task to the PID operation accelerator according to the feature data extracted by the S3, and the PID operation accelerator performs multi-channel parallel PID operation to realize collaborative control of multiple building equipment.

[0013] Preferably, the PID operation adopts an incremental PID algorithm, and the formula is: Δu( k )= K p [E( k )-E( k -1)]+ K i E( k )+ K d [E( k )-2E( k -1)+E( k -2)] Wherein, Δ u ( k ) is a control increment at the moment, k p is a proportional coefficient, K i is an integral coefficient, K d is a differential coefficient, and E( K ) is a deviation at the moment. k k To realize adaptive adjustment of parameters, the PID operation accelerator collects feedback data of the actuator in real time, and dynamically corrects the parameters according to the deviation change trend: when the feedback data deviation continuously increases, the proportional coefficient p is automatically increased to speed up the system response speed; when the system has a steady-state error, the integral coefficient K i is automatically increased to eliminate the steady-state error; when the system response overshoots, the differential coefficient K d is automatically reduced to suppress the overshoot trend; the parameter adjustment result is stored in the storage module in real time to provide data for subsequent adaptive learning. K

[0014] Preferably, the S5 specifically comprises: ​​The general processor reads the filtered data of S3, the PID operation result of S4 and the running state data of each device from the storage module, adopts a weighted fusion method to perform fusion processing on the multi-source data, and generates a high-precision control instruction; The core of the weighted fusion is to dynamically adjust the weight coefficient according to the reliability of the data source, and the specific weight allocation rule is as follows: 1) Analog quantity data weight: determined based on signal quality score, the signal quality score is calculated by corresponding indicators, and the weight coefficient = signal quality score / sum of all data scores; 2) Bus type data weight: determined based on communication success rate, the communication success rate = number of successfully transmitted data packets / total number of transmitted data packets × 100%, and the weight coefficient = communication success rate / sum of all data communication success rates; 3) Digital quantity data weight: determined based on state stability, state stability = stable state duration / total observation time × 100%, and the weight coefficient = state stability / sum of all data state stabilities; After the weight is determined, the weighted sum formula is used for data fusion: fusion result = Σ (data value × corresponding weight coefficient); According to the fusion result, combined with the control logic of the building equipment, specific control instructions are generated and transmitted to the actuator through the bus interface to realize precise control of the building equipment.

[0015] Preferably, S6: adaptive learning and parameter optimization is included. The general processor rolls over every N1 hours to collect the last N2 groups of stored "deviation-control-feedback" triples, trains a lightweight error prediction network, and outputs a hyperparameter correction amount accordingly; The network structure is as follows: a) The input layer has 18 dimensions, of which 10 dimensions are continuous physical characteristics, including current deviation, deviation integral, deviation differential, noise energy ratio before filtering, current filtering window length, normalized Kp / Ki / Kd, signal quality score and device load rate, and 8 dimensions are category features encoded by One-Hot; b) The hidden layer has only one layer with 8 neurons, adopts ReLU activation, and adds Dropout(0.1) and L2 regularization(1e-4) to suppress overfitting; c) The output layer has 5 dimensions, adopts tanh activation, and corresponds to Kalman filter process noise variance correction amount ΔQ, observation noise variance correction amount ΔR, and PID three coefficient correction amounts ΔKp, ΔKi and ΔKd respectively. Each component is linearly mapped to the interval [-0.01, 0.01] or [-0.5, 0.5] according to the engineering allowed step; d) The loss function takes SmoothL1 + weight decay, the optimizer uses AdamW, the learning rate is 3x10 -4 , the batch size is 64, the early stop patience is 20 rounds; e) After the training is completed, the correction amount is added to the original parameter and is limited in amplitude, and is updated to the storage module as the default parameter in the next period.

[0016] Compared with the prior art, the present application has the following remarkable beneficial effects: 1. The self-adaptation and intelligence level are significantly improved Through the multi-modal data intelligent identification of step S1 and the adaptive preprocessing strategy selection of step S2, the differential accurate processing of multi-source heterogeneous data such as analog quantity, digital quantity and bus type is realized, and the problem of insufficient precision caused by the "one-size-fits-all" processing of traditional algorithms is completely solved. Especially the intelligent adaptive learning and parameter optimization mechanism introduced in step S6, through the lightweight neural network and reinforcement learning technology, the control parameters can be dynamically optimized according to the real-time running state of the system, environmental changes and equipment aging degree, so that the algorithm can intelligently adapt to different building scenes (such as the running difference of air conditioning system in summer and winter), different use time periods (such as the change of people flow in daytime and nighttime) and the performance degradation in the whole life cycle of equipment, and the adaptive ability is improved by more than 70%.

[0017] 2. The processing efficiency and real-time performance are greatly optimized Innovatively, the hardware parallel advantages of heterogeneous computing architecture are used to unload the signal filtering and feature extraction tasks of step S3 to a special signal filtering accelerator for parallel processing, and the multi-channel PID operation tasks of step S4 are distributed to a special PID operation accelerator for parallel execution, and the intelligent training scheduler of step S6 is used to dynamically optimize the allocation of computing resources, so that the hardware resources are maximally utilized. Compared with the serial processing mode of traditional general-purpose processors, the data processing delay of the present application is reduced by 55%-65%, reaching the level of 45ms, which fully meets the millisecond-level real-time control requirements of building automation systems, and the overall energy efficiency of the system is improved by more than 25%.

[0018] 3. The control precision and stability are greatly improved The adaptive multi-modal filtering algorithm of step S3 accurately filters out noise for different signal characteristics, the weighted fusion algorithm of step S5 improves the multi-source data fusion accuracy through dynamic weight distribution, the adaptive PID parameter adjustment of step S4 quickly corrects system deviation, and the closed-loop parameter optimization mechanism of step S6 continuously iterates and improves, forming a multi-level and closed-loop precision guarantee system. The measured data shows that the comprehensive control accuracy of building equipment is improved by 70%~80%, such as the temperature control accuracy is improved from ±1.2℃ to ±0.3℃, the humidity control accuracy is improved by more than 60%, at the same time, the system overshoot is reduced by 50%, the steady-state regulation time is shortened by 40%, and the control stability is significantly enhanced.

[0019] 4. System intelligence and self-learning ability breakthrough The introduced intelligent adaptive learning system is an important innovation point of the present application, which has four core intelligent capabilities: - Intelligent fault diagnosis: through data quality detection and abnormal pattern recognition, the fault warning accuracy is more than 98%; - Parameter self-optimization: parameter optimization is automatically performed every 4 hours, and the optimization success rate of more than 10% improvement rate threshold is 96%; - Knowledge transfer ability: support parameter transfer learning between different scenes and different equipment, shorten the adaptation time of new scene by 70%; - Explainable decision: through explainable technologies such as SHAP, provide decision basis for parameter adjustment, and enhance system transparency.

[0020] 5. Energy efficiency and economy are significantly improved Through high-precision control and intelligent optimization algorithm, the energy consumption of building equipment is greatly reduced. In the 10 million square meters of commercial complex, the total energy consumption of air conditioning system is reduced by 18~22%, the energy consumption of lighting system is reduced by 15~20%, and the energy consumption of water supply and drainage system is reduced by 10~15%. According to the annual calculation, a single large building can save energy consumption cost of 300~500 thousand yuan, and the investment return period is shortened to 1.5~2 years. At the same time, the optimization of equipment running time reduces frequent start and stop, and prolongs the service life of equipment by more than 20%.

[0021] 6. System compatibility and expandability are fully enhanced Support BACnet, Modbus, LonWorks and other mainstream industrial bus protocols, compatible with temperature, humidity, pressure, flow and other sensors, and air conditioning, lighting, water supply and drainage, power supply and distribution and other building equipment. The algorithm adopts modular design, which can flexibly adjust parameters according to different building scale and equipment configuration, without large-scale modification of existing building automation system. Support cloud collaboration architecture, can realize centralized optimization management of multiple building groups, system expansion cost is reduced by 40%.

[0022] 7. Maintenance cost and operation efficiency optimization Traditional building automation systems require regular parameter tuning and device calibration by professionals. The present invention realizes self-maintenance and self-optimization of the system through adaptive learning mechanism, reducing maintenance workload by more than 60%. The intelligent monitoring instrument panel and fault warning system enable operation and maintenance personnel to quickly locate problems, shorten the average fault repair time (MTTR) by 50%, and achieve system availability of more than 99.9%.

[0023] 8. Overall improvement of safety and reliability By introducing parameter verification and rollback mechanism, it is ensured that each parameter update is fully verified to prevent system instability caused by incorrect parameters. The data quality detection module effectively filters abnormal data to avoid the "garbage in, garbage out" problem. The system has multiple backup and recovery mechanisms to automatically recover to a safe state when training fails or verification fails, ensuring continuous and stable operation of the system.

[0024] In summary, the present invention realizes the overall improvement of building automation system in adaptive ability, processing efficiency, control accuracy, energy efficiency, system compatibility and operation economy through the organic combination of heterogeneous computing architecture, multi-modal data processing, adaptive control algorithm and intelligent learning system, providing a complete technical solution for efficient, accurate and reliable operation of intelligent buildings. It has significant technical advancement and economic practicality. BRIEF DESCRIPTION OF DRAWINGS

[0025] The present invention will be further described in detail below in conjunction with the accompanying drawings: Figure 1 : Adaptive multi-modal data processing algorithm flowchart; Figure 2 : Data classification and mode identification schematic diagram; Figure 3 : Adaptive filtering strategy selection logic diagram; Figure 4 : Data fusion and instruction generation flowchart; Figure 5 : Adaptive learning and optimization algorithm complete flowchart; Figure 6 : Data acquisition and preprocessing flowchart; Figure 7: Intelligent training strategy selection flowchart; Figure 8: Neural network structure diagram; Figure 9: Adaptive limiting algorithm flowchart; Figure 10: Performance monitoring and feedback algorithm flowchart. DETAILED DESCRIPTION

[0026] To achieve the above object, the technical scheme adopted by the present application is as follows: the algorithm runs in a building automatic control real-time data processing system based on a heterogeneous computing architecture, the heterogeneous computing architecture at least includes a general processor (CPU), a signal filtering accelerator, a PID operation accelerator and a storage module, as shown in Figures 1-10 The algorithm specifically includes the following steps: Step S1: multi-source data classification and modality recognition After the multi-source data in the building automatic control system is collected by sensors, actuators and bus interfaces, it is transmitted to the general processor of the heterogeneous computing architecture, and the data classification and modality recognition process is performed by the general processor. First, according to the source and transmission characteristics of the data, it is divided into three categories: analog quantity, digital quantity and bus type data. Analog quantity data is derived from temperature sensors, humidity sensors, pressure sensors and other analog signal acquisition devices, and exists in the form of continuously changing voltage or current signals; digital quantity data is derived from device switches, buttons, counters and other digital signal devices, and exists in the form of discrete 0 / 1 or specific coding; bus type data is derived from devices connected through industrial buses (such as BACnet, Modbus), and is transmitted in the form of frame structure.

[0027] On the basis of classification, modality recognition is performed on each type of data: for analog quantity data, the periodic characteristics, mutation frequency and variance value of the signal are calculated to identify it as a periodic signal (such as the return air temperature signal of an air conditioning system, the period is usually 10-15 minutes), a mutation signal (such as the pressure mutation signal when the fan starts and stops), and a steady-state signal (such as the lighting current signal when the room is unoccupied); for digital quantity data, the duration, frequency of change and coding rules of the signal are analyzed to identify it as an on-off quantity (such as the on-off state of a lamp), a pulse count (such as the pulse signal of an electric meter), and a state quantity (such as the running / fault state of a device); for bus type data, the header identifier, protocol field and check bit of the data frame are analyzed to identify the protocol type (such as BACnet IP, Modbus RTU) and data frame structure (such as data length, field meaning), and the effective data in the frame is extracted. The recognition result is stored in the storage module at the same time, providing a basis for subsequent processing strategy selection.

[0028] Step S2: adaptive pre-processing strategy selection The general processor dynamically matches the pre-processing strategy for different types of data according to the classification and modality recognition result of step S1, and the pre-processing process is completed by the general processor and the storage module in cooperation, and the specific strategy is as follows: 1. Analog data preprocessing: In view of the characteristics of analog data being susceptible to noise interference and having nonlinear error, a combined strategy of "linearization calibration-outlier rejection-dimension normalization" is adopted. Linearization calibration converts the collected voltage / current signal into actual physical quantity (e.g., converts 0~5V voltage signal into 0~100℃ temperature value) by querying the sensor calibration curve, and corrects the nonlinear error of the sensor; Outlier rejection uses 3 σ criteria to calculate the mean and standard deviation of the data, and data outside the range of [mean-3 σ , mean+3 σ ] is judged as an outlier and is removed to avoid the influence of extreme data on subsequent processing; Dimension normalization converts different physical quantities (e.g., temperature ℃, humidity %RH) to the [0,1] interval, formula: normalized value=(original value-minimum value) / (maximum value-minimum value), which provides a unified dimension for subsequent feature extraction and data fusion.

[0029] 2. Digital data preprocessing: In view of the problems of digital data prone to jitter and state coding not unified, the strategy of "de-jitter processing-state coding" is adopted. De-jitter processing sets a time window (e.g., 50ms), if the signal state continuously remains the same within the window, it is confirmed as a valid state, otherwise it is considered as jitter and ignored; State coding converts the discrete states of different devices into unified binary coding (e.g., "running" state is coded as 10, "fault" state is coded as 01, and "standby" state is coded as 00), which improves the standardization of data processing.

[0030] 3. Bus type data preprocessing: In view of the characteristics of bus type data having protocol differences and data transmission being prone to errors, the strategy of "protocol analysis-data validity check" is adopted. According to the protocol type identified in step S1, the corresponding protocol analysis module is called to extract the device address, data identifier, value and other key information in the data frame; Data validity check checks the check bit (such as CRC check) at the end of the frame to determine whether there is packet loss or error code during data transmission. If the check fails, retransmission is requested to ensure data integrity.

[0031] Step S3: Parallel filtering and feature extraction The preprocessed data of various types is dispatched by the general processor to the signal filtering accelerator of the heterogeneous computing architecture to realize parallel filtering and feature extraction, fully utilizing the parallel processing capability of the hardware accelerator. The signal filtering accelerator adaptively selects the filtering method according to the analog data mode identified in step S1, as follows: 1. Periodic signal filtering: Kalman filtering algorithm is adopted to dynamically estimate the true value of the signal by establishing the state equation and observation equation of the signal. Taking the air return temperature signal of the air conditioner as an example, the state equation is set as: X(k )=A*X( k -1)+B* u ( k )+ w ( k ) Where X( k )for k The temperature state at time t, A is the state transition matrix, and B is the control matrix. u ( k ) is the control quantity. w ( k ) represents process noise; the observation equation is set as Z( k )=H*X( k )+ v ( k ), where Z( k )for k The observations at time t, where H is the observation matrix. v ( k () represents observation noise. Through a prediction-update iterative process, random noise in periodic signals is effectively filtered out.

[0032] 2. Abrupt Signal Filtering: An adaptive threshold filtering algorithm is used to dynamically adjust the threshold based on the historical fluctuations of the signal. First, the mean of the first N sampling points is calculated. μ With respect to the standard deviation σ, set the threshold to μ ± k σ( k To adjust the threshold coefficient (usually 2-3), when the current sampling point exceeds the threshold range, it is determined to be a sudden signal and retained; otherwise, it is considered noise and filtered out. This method can effectively preserve the peak characteristics of sudden signals and avoid the misjudgment of sudden signals by traditional fixed threshold filtering.

[0033] 3. Steady-state signal filtering: A moving average filtering algorithm is used, with a sliding window of length M (M is adjusted according to the steady-state characteristics of the signal, usually 5~10). The average value of the sampled data within the window is taken as the filtering result at the current time. The formula is: Y( k )=(X( k )+X( k -1)+…+X( k -M+1)) / M, effectively smooths high-frequency noise in steady-state signals.

[0034] After filtering, the signal filtering accelerator synchronously extracts key features from various types of data: for analog data, it extracts features such as mean, variance, peak value, and period; for digital data, it extracts features such as state duration and switching frequency; and for bus-type data, it extracts features such as data update frequency and transmission delay. The extracted feature data is stored in the storage module to support subsequent control calculations and data fusion.

[0035] Step S4: Execution of the multi-task cooperative control algorithm Based on the feature data extracted in step S3, the general-purpose processor assigns the control task to the PID accelerator. The PID accelerator then performs multi-channel parallel PID calculations to achieve coordinated control of multiple building devices. The PID accelerator supports 8 to 32 parallel channels, with each channel corresponding to a controlled object (such as an air conditioning unit or a lighting circuit). The channels operate independently, significantly improving control efficiency.

[0036] PID calculation uses the incremental PID algorithm, and the formula is: Δu( k )= K p [E( k )-E( k -1)]+ K i E( k )+ K d [E( k )-2E( k -1)+E( k -2)] Where, Δ u ( k )for k Control increment at any time, K p This is the proportionality coefficient. K i The integral coefficient is... K d E( is the differential coefficient) k )for k The deviation in time (the difference between the set value and the feedback value).

[0037] To achieve adaptive parameter adjustment, the PID calculation accelerator collects feedback data from the actuator in real time (such as valve opening and fan speed) and dynamically corrects the parameters based on the trend of deviation changes: when the feedback data deviation continues to increase (such as after 3 consecutive sampling cycles E( k )>E( k -1)>E( k When -2), the scaling factor is automatically increased. K p (Increase the ratio by 10%~20%) to accelerate the system response speed; when the system exhibits steady-state error (such as deviation E( k When the absolute value of the integral remains within the range of 0.5~1℃ for 5 consecutive sampling periods, the integral coefficient is automatically increased. K i(5%~15% increase), eliminate steady-state error; when the system response overshoots (such as the feedback value exceeds the set value by more than 5%), automatically reduce the derivative coefficient K d (20%~30% reduction), suppress overshoot tendency. The parameter adjustment results are stored in real time to the storage module to provide data for subsequent adaptive learning.

[0038] Step S5: Data fusion and output instruction generation The general processor reads the filtered data of step S3, the PID operation results of step S4, and the running state data of each device from the storage module, adopts a weighted fusion method to perform fusion processing on the multi-source data, and generates a high-precision control instruction. The core of weighted fusion is to dynamically adjust the weight coefficient according to the reliability of the data source. The specific weight allocation rules are as follows: 1. Analog data weight: determined based on signal quality score, signal quality score is calculated by noise intensity, sampling frequency, calibration error, etc. (full score is 10 points), weight coefficient = signal quality score / sum of all data scores. For example, the temperature sensor signal quality score is 9 points, the humidity sensor score is 8 points, and the bus type data score is 7 points, then the temperature data weight is 9 / (9+8+7)=0.375, the humidity data is 0.333, and the bus data is 0.292.

[0039] 2. Bus type data weight: determined based on communication success rate, communication success rate = number of successfully transmitted data packets / total number of transmitted data packets x 100%, weight coefficient = communication success rate / sum of all data communication success rates. If the communication success rate of a certain bus device is 95% and that of another device is 90%, the weights are 95 / (95+90)=0.514 and 0.486 respectively.

[0040] 3. Digital data weight: determined based on state stability, state stability = stable state duration / total observation time x 100%, weight coefficient = state stability / sum of all data state stability. If the stable duration of a certain switch state accounts for 98%, and the stable duration of another state accounts for 92%, the weights are 98 / (98+92)=0.516 and 0.484 respectively.

[0041] After determining the weights, the weighted sum formula is used for data fusion: fusion result = Σ(data value x corresponding weight coefficient). According to the fusion result, combined with the control logic of the building equipment (such as the temperature control logic of the air conditioning system, the human body sensing logic of the lighting system), specific control instructions (such as "air conditioning unit compressor frequency is raised to 50Hz" "lighting circuit 1 is turned off") are generated, and transmitted to the actuators through the bus interface, to realize precise control of the building equipment.

[0042] Step S6: Adaptive Learning and Parameter Optimization To continuously optimize algorithm performance, the general-purpose processor continuously collects the most recently stored 1000 sets of "bias-control-feedback" triples every 4 hours to train a lightweight error prediction network, and outputs hyperparameter corrections accordingly. The network structure is as follows: a) The input layer has 18 dimensions, of which 10 dimensions are continuous physical features (including current deviation, deviation integral, deviation derivative, noise-energy ratio before filtering, current filtering window length, normalized Kp / Ki / Kd and signal quality score), and 8 dimensions are One-Hot encoded categorical features (modal type, data type, bus protocol). b) The hidden layer consists of only 1 layer with 8 neurons, using ReLU activation, and Dropout (0.1) and L2 regularization (1e-4) are added to suppress overfitting; c) The output layer is 5-dimensional and uses tanh activation. It corresponds to the noise variance correction ΔQ, the observation noise variance correction ΔR, and the PID three coefficient corrections ΔKp, ΔKi, and ΔKd, respectively. Each component is linearly mapped to the interval [-0.01, 0.01] or [-0.5, 0.5] according to the engineering allowable step size. d) The loss function is SmoothL1 (prediction bias and actual bias under network output correction) + weight decay. The optimizer is AdamW with a learning rate of 3×10⁻⁶. -4 Batch size 64, early stop patience 20 rounds; e) After training is complete, the correction is added to the original parameters and the amplitude is limited, then updated to the storage module as the default parameters for the next cycle.

[0043] The complete flowchart of the algorithm is as follows: Figure 5 As shown.

[0044] Detailed steps of the algorithm: 6.1 Data Acquisition and Preprocessing Module 1. The system automatically collects the "deviation-control-feedback" triplet data for the most recent 4 hours; 2. Perform data quality checks and remove abnormal data; 3. Perform interpolation to repair missing data; 4. Extract an 18-dimensional feature vector (10-dimensional continuous features + 8-dimensional categorical features). 5. Normalize and encode the features.

[0045] Technical features: 1. Use sliding window statistics for feature normalization to avoid data leakage; 2. Implement One-Hot encoding to process category features; 3. Built-in data quality scoring mechanism (0~1 point), only retaining data with a score ≥0.7.

[0046] The data acquisition and preprocessing process is shown in Figure 6.

[0047] 6.2 Intelligent Training Strategy Selection Module 1. Analyze the current status and historical data of the system; 2. Based on indicators such as error trend, data quality, and time since the last training; 3. Dynamically select one of four training strategies: 1) Full training: When the error increases, train all network parameters; 2) Incremental training: When the error is stable, only the output layer is trained; 3) Transfer learning: When the error decreases, fine-tune the pre-trained model; 4) Warm start training: Use the base model when data is insufficient.

[0048] Technical features: 1. Adaptive training frequency adjustment (1~12 hours); 2. Supports online learning and offline batch learning; 3. Built-in early stopping mechanism to prevent overfitting.

[0049] Figure 7. Intelligent training strategy selection process.

[0050] 6.3 Neural Network Optimization Module 1. Construct an 18-8-5 three-layer neural network structure; 2. Input layer: 18-dimensional feature vector; 3. Hidden layer: 8 neurons, ReLU activation, Dropout (0.1) regularization; 4. Output layer: 5-dimensional parameter correction, tanh activation; 5. Train using the SmoothL1 loss function and the AdamW optimizer.

[0051] The neural network structure is shown in Figure 8.

[0052] Table 1: Detailed Description of Network Structure

[0053] Table 2: Feature Mapping Relationship Table - Continuous Features (10 Dimensions)

[0054] Table 3: Feature Mapping Relationship Table – Categorical Features (8 Dimensions)

[0055] Table 4: Output Parameter Description Table

[0056] Table 5: Training Parameter Configuration Table

[0057] 6.4 Parameter Correction and Verification Module 1. Adaptively limit the correction amount of the neural network output; 2. Dynamically adjust the amplitude limit range according to the system status; 3. Add the correction amount to the current parameter; 4. Test the effect of the new parameters on the validation dataset; 5. Update system parameters only when the improvement rate is ≥10%.

[0058] The adaptive limiting algorithm is shown in Figure 9.

[0059] 6.5 Performance Monitoring and Feedback Module 1. Real-time monitoring and optimization of algorithm execution performance; 2. Record detailed logs of all optimization operations; 3. Generate a visual performance report; 4. Adjust training strategies and frequency based on historical performance; 5. Provide fault diagnosis and recovery mechanisms.

[0060] Monitoring indicator system: 1. Control accuracy indicators: average error, maximum error, steady-state error; 2. System stability indicators: oscillation frequency, overshoot, and settling time; 3. Algorithm performance metrics: training time, memory usage, convergence speed; 4. Energy efficiency indicators: energy consumption reduction rate, equipment operating time, peak demand.

[0061] The flowchart of the performance monitoring and feedback algorithm is shown in Figure 10.

[0062] Specific applications: To make the technical solution of this invention clearer and more explicit, the following detailed description is based on an air conditioning automatic control system of a large commercial complex as a specific embodiment. This commercial complex has a building area of ​​100,000 square meters and is equipped with an intelligent building automatic control system. The system architecture adopts a heterogeneous computing platform, and the specific hardware configuration is as follows: System hardware configuration: 1. General-purpose processor (CPU): Intel Core i7-12700, clock speed 2.1~4.9GHz, 12 cores and 20 threads, used for system scheduling and algorithm coordination; 2. Signal filtering accelerator: Xilinx Kintex-7 XC7K325T FPGA, with 256 built-in DSP slices, dedicated to parallel filtering operations; 3. PID Calculation Accelerator: TI TMS320C6748 DSP, supporting 32-channel parallel PID calculation, with a main frequency of 456MHz; 4. Storage module: 8GB DDR4 3200MHz memory, 512GB NVMe SSD solid-state drive; 5. Data acquisition interface: Supports multiple protocols including BACnet / IP, Modbus TCP / RTU, and LonWorks. System software configuration: 1. Operating System: Ubuntu 20.04 LTS real-time kernel version; 2. Algorithm framework: PyTorch 1.12.0 (training phase), LibTorch 1.12.0 (inference phase); 3. Communication protocol stack: BACnet Stack 1.3, Modbus Library 2.0; 4. Monitoring System: A custom monitoring dashboard based on Grafana.

[0063] System access devices: 1. Analog signal equipment: 200 temperature sensors (PT1000, accuracy ±0.1℃), 50 humidity sensors (accuracy ±2%RH), and 30 pressure sensors (accuracy ±0.5%). 2. Digital signal devices: 20 start / stop switches for air conditioning units, 50 fault alarm contacts, and 100 lighting circuit switches; 3. Bus equipment: 30 BACnet / IP air conditioning terminal controllers and 20 Modbus RTU smart meters; 4. Actuators: 15 frequency converters (control accuracy 0.1Hz), 40 electric valves (control accuracy 0.5%). Example

[0064] Step S1: Multi-source data classification and modality recognition After the multi-source data of the air conditioning automatic control system is collected, it is transmitted to the Intel Core i7-12700 CPU. The CPU first classifies the data: the return air temperature, supply water temperature and outdoor temperature of the 10 air conditioning units are classified as analog data; the start / stop switch and fault alarm signals of the air conditioning units are classified as digital data; and the fan speed and valve opening signals of the terminal controller transmitted through the BACnet IP bus are classified as bus data.

[0065] Modal identification is then performed: by calculating the signal period, the return air temperature (cycle approximately 12 minutes) and supply water temperature (cycle approximately 15 minutes) are identified as periodic signals; the supply water pressure signal during air conditioning unit start-up and shutdown (pressure change amplitude > 0.5 MPa at start-up and shutdown) is identified as abrupt signal; the outdoor temperature (change < 0.3℃ within 1 hour when there is no human activity and no weather changes) is identified as a steady-state signal; the start-up and shutdown status of the air conditioning unit is identified as a switching quantity, and the duration signal of the fault alarm is identified as a status quantity; the bus data is parsed into the BACnet IP protocol, and the data frame structure is "header (16 bytes) + device address (4 bytes) + data identifier (2 bytes) + value (4 bytes) + check bit (2 bytes)". The identification results are stored in an 8GB DDR4 storage module.

[0066] Step S2: Adaptive preprocessing strategy selection The CPU performs preprocessing based on the identification results: for analog return air temperature data, it queries the sensor calibration curve (0-5V corresponds to 0~100℃) for linearization calibration, using 3... σ The criteria remove outliers (such as an outlier of 200℃ caused by sensor failure) and normalize the temperature values ​​to the [0,1] range; for the digital data of the air conditioning unit start / stop switch, a 50ms time window is set for de-jitter processing, and "run" is encoded as 10 and "stop" is encoded as 01; for BACnet IP bus data, the BACnet protocol parsing module is called to extract the device address and valve opening value, and the validity of the data is confirmed by CRC check (retransmission is requested if the check fails).

[0067] Step S3: Parallel Filtering and Feature Extraction The CPU schedules a Xilinx Kintex-7 FPGA signal filtering accelerator to perform parallel processing on the preprocessed data: Kalman filtering is used for periodic signals such as return air temperature and supply water temperature, with a state transition matrix A=0.98, observation matrix H=1, process noise variance Q=0.01, and observation noise variance R=0.05; adaptive threshold filtering is used for sudden changes in supply water pressure. N =10 sampling points, k =2.5, threshold is μ ±2.5σ The outdoor temperature steady-state signal is filtered using a moving average with a window length of M=8. After filtering, the mean return air temperature (e.g., 25℃), variance (e.g., 0.8), and period (12 minutes) are extracted. The duration of the switch status (e.g., running for 2 hours) is extracted, as well as the update frequency of the bus data (e.g., updated every 5 seconds). The feature data is then stored in the storage module.

[0068] Step S4: Execution of the multi-task cooperative control algorithm The CPU assigns the control tasks of 10 air conditioning units to the TI TMS320C6748 DSP PID accelerator, enabling 10 parallel channels to perform incremental PID calculations, with the target return air temperature set at 26℃. When the actual return air temperature is 28℃, and the deviation continues to increase (deviations of 2℃, 2.1℃, and 2.2℃ for three consecutive sampling periods), the PID accelerator automatically adjusts... Kp Increase Ki from 5 to 5.8; when the temperature stabilizes at 26.5℃ and a steady-state error of 0.5℃ occurs (lasting for 5 sampling periods), increase Ki from 0.1 to 0.12; when the temperature suddenly drops to 25℃ and an overshoot of 0.5℃ occurs, increase Ki from 5.8 to 5.8. Kd The value decreases from 1 to 0.7. The calculation results in control increment signals such as "compressor frequency increased to 55Hz" and "electronic expansion valve opening increased to 80%".

[0069] Step S5: Data fusion and output instruction generation The CPU reads the filtered return air temperature (28℃, signal quality score 9), the valve opening transmitted via the bus (70%, communication success rate 96%), and the digital operating status (stable duration percentage 98%). It calculates the weights as follows: temperature data weight = 9 / (9+9.6+9.8) = 0.302, valve opening weight = 9.6 / (9+9.6+9.8) = 0.322, and operating status weight = 9.8 / (9+9.6+9.8) = 0.376. After weighted fusion, a comprehensive control index is obtained. Combined with the air conditioning system control logic (increasing compressor frequency when return air temperature > 26℃), a control command is generated: "Increase the compressor frequency of air conditioning unit 1 to 55Hz, adjust the electronic expansion valve opening to 80%." This command is transmitted to the actuator via the BACnet IP bus.

[0070] Step S6: Adaptive Learning and Parameter Optimization Every 4 hours, the process is rolled over: The CPU automatically extracts the latest 1000 sets of local "bias-control-feedback" samples (approximately 80 MB of storage) to train an 18-8-5 lightweight error prediction network. The input is 18 dimensions (10 continuous physical features + 8 One-Hot class features), with 8 hidden layers (ReLU + Dropout 0.1 + L2 = 1e-4), and the output is a 5-dimensional tanh mapping: ΔQ, ΔR, ΔKp, ΔKi, ΔKd, with amplitudes limited to [-0.01, 0.01] or [-0.5, 0.5] respectively. The loss function is SmoothL1, AdamW lr = 3 × 10⁻¹⁰. -4 The batch size was set to 64, and the early stopping patience was set to 20. Upon completion of training, the correction was added to the original parameters and the amplitude was limited. The data was then written back to the storage module as the default value for the next cycle. This "small step, fast run" rolling optimization, in a real-world test on a 100,000㎡ commercial building, further reduced the steady-state temperature error from ±0.5℃ to ±0.3℃ within 12 hours, without exhibiting parameter oscillations.

[0071] Example Effect Verification Statistics on the continuous operation of the air conditioning control system (100,000 m², 10 screw chillers, 200+ sensors) in the same commercial complex for 30 days: 1. Data processing latency decreased from 100 ms to 45 ms (↓55%); 2. The return air temperature control accuracy has been improved from ±1.2 ℃ to ±0.3 ℃ (↑75%). 3. Total unit energy consumption reduced by 20% (high precision reduces frequent start-stop cycles); 4. The effective data transmission rate of the bus has increased from 92% to 98%, significantly enhancing system stability.

[0072] The air conditioning automatic control system of the commercial complex operated continuously for 30 days, and the statistical performance data are as follows: Table 6: Control Performance Indicators

[0073] Table 7: Energy Efficiency Indicators

[0074] Table 8: Algorithm Performance Metrics

[0075] Table 9: System Reliability Indicators

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. In practical applications, algorithm parameters (such as learning cycle, filter window size, PID parameter adjustment ratio, etc.) can be adjusted according to the building scale, number of devices, and specific configuration of heterogeneous computing architecture. All modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of the present invention.

Claims

1. A self-adapting multi-modal building equipment data processing method for heterogeneous computing architecture, characterized in that, Comprising the following steps: S1: multi-source data classification and modal recognition; S2: adaptive preprocessing strategy selection; S3: parallel filtering and feature extraction; S4: multi-task collaborative control algorithm execution; S5: data fusion and output instruction generation.

2. The adaptive multi-modal building equipment data processing method for heterogeneous computing architecture according to claim 1, characterized in that: S1 specifically comprises: According to the source and transmission characteristics of the data, it is divided into three categories: analog quantity, digital quantity and bus type data; On the basis of classification, modal recognition is carried out on each type of data: for analog quantity data, the period characteristics, mutation frequency and variance value of the signal are calculated to identify it as a periodic signal, a mutation signal or a steady-state signal; For digital quantity data, the duration, frequency of change and coding rules of the signal are analyzed to identify it as a switch quantity, pulse count or state quantity; For bus type data, the protocol type and data frame structure are identified by analyzing the header identification, protocol field and check bit of the data frame, and the valid data in the frame is extracted; The identification result is stored in the storage module simultaneously to provide a basis for subsequent processing strategy selection.

3. The adaptive multi-modal building equipment data processing method for heterogeneous computing architecture according to claim 1, characterized in that: S2 specifically comprises: According to the classification and modal recognition result of S1, different types of data are dynamically matched with preprocessing strategies, and the preprocessing process is completed by the general processor and the storage module in cooperation.

4. The adaptive multi-modal building equipment data processing method for heterogeneous computing architecture according to claim 3, characterized in that: The preprocessing strategy is as follows: 1) Analog quantity data preprocessing: According to the characteristics of analog quantity data being easily disturbed by noise and having nonlinear errors, a combined strategy of "linearization calibration- outlier rejection- dimensionless normalization" is adopted; Linearization calibration converts the collected voltage / current signal into actual physical quantity by querying the sensor calibration curve, correcting the non-linear error of the sensor; Outlier rejection adopts 3 σ criteria, calculates the mean and standard deviation of the data, and determines the data beyond the range of [mean-3 σ , mean+3 σ ] as outliers and rejects them, avoiding the influence of extreme data on subsequent processing; Dimensionless normalization converts different physical quantities to the [0, 1] interval, and the formula is: normalized value = (original value - minimum value) / (maximum value - minimum value), which provides a unified dimension for subsequent feature extraction and data fusion; 2) Digital quantity data preprocessing: In view of the problems of digital quantity data being prone to jitter and state coding being not unified, a "debouncing processing-state coding" strategy is adopted; Debouncing processing sets a time window, if the signal state remains consistent within the window, it is confirmed as a valid state, otherwise it is considered as jitter and ignored; State coding converts the discrete states of different devices into a unified binary code, improving the standardization of data processing; 3) Bus type data preprocessing: In view of the characteristics of bus type data having protocol differences and data transmission being prone to errors, a "protocol analysis-data validity check" strategy is adopted; Protocol analysis calls the corresponding protocol analysis module according to the protocol type identified in S1 to extract the key information in the data frame; Data validity check judges whether the data has packet loss or error code in the transmission process by checking the check bit at the end of the frame, if the check fails, it requests retransmission to ensure the integrity of the data.

5. The adaptive multi-modal building equipment data processing method for heterogeneous computing architecture according to claim 1, wherein the S3 specifically comprises: After the S2 preprocessing, the various types of data are dispatched to the signal filtering accelerator of the heterogeneous computing architecture by the general processor to realize parallel filtering and feature extraction; the signal filtering accelerator adaptively selects a filtering method according to the analog data modality identified in the S1.

6. The adaptive multi-modal building equipment data processing method for heterogeneous computing architecture according to claim 5, wherein the filtering method comprises: 1) periodic signal filtering: a Kalman filtering algorithm is used to dynamically estimate the true value of the signal by establishing the state equation and observation equation of the signal; 2) sudden change signal filtering: an adaptive threshold filtering algorithm is used to dynamically adjust the threshold according to the historical fluctuation of the signal; After filtering, the signal filtering accelerator synchronously extracts the key features of various types of data; the extracted feature data are stored in the storage module to support subsequent control operation and data fusion.

7. The adaptive multi-modal building equipment data processing method for heterogeneous computing architecture according to claim 1, wherein the S4 specifically comprises: 3) Steady-state signal filtering: A moving average filtering algorithm is used, with a sliding window of length M. The average value of the sampled data within the window is taken as the filtering result at the current time. The formula is: Y( k )=(X( k )+X( k -1)+…+X( k -M+1)) / M; The general processor assigns a control task to the PID operation accelerator according to the feature data extracted in the S3, and the PID operation accelerator performs multi-channel parallel PID operation to realize cooperative control of multiple building equipment.

8. The adaptive multi-modal building equipment data processing method for heterogeneous computing architecture according to claim 7, wherein the PID operation adopts an incremental PID algorithm, and the formula is:

9. The adaptive multi-modal building equipment data processing method for heterogeneous computing architecture according to claim 1, wherein the S5 specifically comprises: The general processor reads the filtered data of the S3, the PID operation result of the S4, and the running state data of each device from the storage module, and adopts a weighted fusion method to fuse the multi-source data to generate a high-precision control instruction; The core of the weighted fusion is to dynamically adjust the weight coefficient according to the reliability of the data source, and the specific weight allocation rules are as follows: 1) analog data weight: determined based on signal quality score, the signal quality score is calculated by corresponding indicators, and the weight coefficient = signal quality score / sum of all data scores; Δu( k )= K p [E( k )-E( k -1)]+ K i E( k )+ K d [E( k )-2E( k -1)+E( k -2)] Where, Δ u ( k )for k Control increment at any time, K p This is the proportionality coefficient. K i The integral coefficient is... K d E( is the differential coefficient) k )for k Timing deviation; To realize the self-adaptive adjustment of parameters, the PID operation accelerator collects the feedback data of the actuator in real time and dynamically corrects the parameters according to the change trend of the deviation: when the deviation of the feedback data continuously increases, the proportional coefficient is automatically increased K p , the system response speed is accelerated; when the system has a steady-state error, the integral coefficient is automatically increased K i , the steady-state error is eliminated; when the system response overshoots, the differential coefficient is automatically reduced K d , the overshoot trend is suppressed; the results of parameter adjustment are stored in the storage module in real time to provide data for subsequent self-adaptive learning. 2) bus type data weight: determined based on communication success rate, the communication success rate = number of successfully transmitted data packets / total number of transmitted data packets × 100%, and the weight coefficient = communication success rate / sum of all data communication success rates; 3) digital quantity data weight: determined based on state stability, the state stability = stable state duration / total observation time × 100%, and the weight coefficient = state stability / sum of all data state stabilities; After the weight is determined, the weighted sum formula is used for data fusion: fusion result = Σ (data value × corresponding weight coefficient). ​ ​ ​ ​ ​ According to the fusion result, combined with the control logic of the building equipment, specific control instructions are generated and transmitted to the actuators through the bus interface to realize precise control of the building equipment.

10. The adaptive multi-modal building equipment data processing method for heterogeneous computing architecture according to claim 1, characterized in that, S6: adaptive learning and parameter optimization; The general processor rolls up the last N2 groups of stored "deviation-control-feedback" triplets every N1 hours, trains a lightweight error prediction network, and outputs a super parameter correction amount accordingly; The network structure is as follows: a) The input layer has a total of 18 dimensions, of which 10 dimensions are continuous physical characteristics, including current deviation, deviation integral, deviation differential, noise energy ratio before filtering, current filtering window length, normalized Kp / Ki / Kd, signal quality score, and equipment load rate, and 8 dimensions are category features encoded by One-Hot; b) The hidden layer has only one layer with 8 neurons, adopts ReLU activation, and adds Dropout(0.1) and L2 regularization(1e-4) to suppress overfitting; c) The output layer has 5 dimensions, adopts tanh activation, and corresponds to Kalman filter process noise variance correction amount ΔQ, observation noise variance correction amount ΔR, and PID three coefficient correction amounts ΔKp, ΔKi, and ΔKd, respectively. Each component is linearly mapped to the interval [-0.01, 0.01] or [-0.5, 0.5] according to the engineering allowed step size; d) The loss function takes SmoothL1 + weight decay, the optimizer takes AdamW, the learning rate is 3x10 -4 , the batch size is 64, and the early stop patience is 20 rounds. e) After training is completed, the correction amount is added to the original parameter and limited in amplitude, and is updated to the storage module as the default parameter for the next period.

Images