AI algorithm engine system applied to Internet of Things platform

Abstract

The invention relates to the technical field of intelligent Internet of Things, in particular to an AI algorithm engine system applied to an Internet of Things platform, which comprises the following steps: a data verification module, a signal noise reduction module, a feature extraction module, a model construction module and a resource scheduling module. According to the method, the key rotation rhythm is dynamically adjusted through data fluctuation characteristics and node synchronous offset, the integrity and safety of data transmission are enhanced, filtering parameters are set through linkage of frequency distribution and energy trend, the de-noising adaptability of non-stationary signals is improved, the sampling frequency is adjusted according to the change rate, and the de-noising efficiency of the non-stationary signals is improved. The method improves the extraction precision of the operation features of the multi-source equipment, dynamically distributes the training frequency in combination with the feature weight and the learning convergence condition, optimizes the training resource distribution and convergence efficiency, enhances the response elasticity of the model structure to the task pressure, matches the scheduling sequence according to the task urgency and the node pressure bearing capability, and improves the efficiency of the model structure. And the coordination and the real-time performance of resource utilization in a concurrent task environment are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent Internet of Things technology, and in particular to an AI algorithm engine system applied to an Internet of Things platform. Background Art

[0002] The field of intelligent Internet of Things technology includes multiple technical branches such as data collection and transmission, edge computing, machine learning model deployment, real-time reasoning optimization, and device collaborative control. This field focuses on the deep integration of artificial intelligence algorithms and IoT terminal devices. By building an end-to-end perception, analysis, and decision-making closed-loop architecture, it upgrades the traditional IoT's one-way data transmission mode to an intelligent system with autonomous decision-making capabilities. Its technical system covers multiple key links including real-time processing of sensor data, lightweight model training, distributed computing resource scheduling, and device control strategy generation, and is applied to multiple scenarios such as industrial equipment status prediction, dynamic monitoring of urban environment, and smart home automation control.

[0003] The AI algorithm engine system applied to the IoT platform addresses the efficient processing and intelligent decision-making needs of multi-source heterogeneous data in the IoT platform, achieving end-to-end conversion of data into control instructions through a combination of multiple technologies. The patent covers technical matters including data denoising and dimensionality compression methods based on wavelet transform and principal component analysis, a time series feature parameter extraction process based on a sliding window mechanism, a model parameter optimization process using long short-term memory networks and stochastic gradient descent, a computational redundancy elimination strategy combining model quantization and graph optimization, a parallel computing framework based on priority strategies and dynamic task decomposition, a control strategy generation mechanism utilizing a rule engine and reinforcement learning, a data security verification method based on a combination of hash checksums and symmetric encryption public key infrastructure, and a system log management process integrating anomaly detection algorithms. By optimizing algorithms and system architecture, the IoT platform can autonomously complete closed-loop responses from data acquisition to device control. This includes shortening device anomaly detection time in industrial scenarios, improving the accuracy of multi-sensor data fusion in urban environmental monitoring, and dynamically optimizing energy consumption in smart home scenarios, ensuring that data transmission and model inference processes meet reliability and security requirements.

[0004] Traditional IoT platform technology mainly relies on fixed parameter settings and general strategy execution for IoT data processing. For example, key rotation is executed at a fixed period and cannot be adjusted in time according to the sensor status. There is a risk of data leakage in high-frequency mutation data scenarios. The filter interval is statically set in the signal processing process, and the frequency response lag leads to residual high-frequency noise or weakening of low-frequency signals, affecting the effectiveness of subsequent analysis. Fixed-frequency acquisition is used without adjusting the change rate of different channels, which can easily cause uneven feature expression or redundant data accumulation. All feature parameters are learned equally in model training, and low-value features occupy training resources, resulting in extended learning cycles and slow convergence. The inference path structure is solidified after the model is built, lacking flexible switching capabilities when facing changes in task scale. The resource scheduling stage is only statically allocated based on task length or node idleness, which cannot cope with the different demands of concurrent tasks, resulting in delays in high-priority task processing and reduced overall system scheduling efficiency. Summary of the Invention

[0005] In order to solve the technical problems existing in the prior art, the embodiment of the present invention provides an AI algorithm engine system applied to the Internet of Things platform. The technical solution is as follows:

[0006] On the one hand, an AI algorithm engine system applied to the Internet of Things platform is provided, which includes:

[0007] The data verification module obtains the monitoring records of the sensor nodes, performs statistical analysis on the data fluctuations in each time period, extracts the data fluctuation characteristic indicators, and adjusts the key rotation cycle based on the operation frequency of the transmission node and the data synchronization offset value to generate a key update sequence list;

[0008] The signal noise reduction module calls the key update time sequence table, collects the data frequency distribution and corresponding signal energy value in the sliding window, adjusts the frequency segment range and filter threshold setting in the signal processing process according to the change trend of the high-frequency energy value and the moving direction of the frequency concentration area, and obtains the denoised signal data;

[0009] The feature extraction module uses the denoised signal data and the data change rate to adjust the sampling frequency parameters of the data, extracts the data features of the operating parameters of various IoT devices, and generates a feature extraction sequence;

[0010] The model construction module extracts the variation range and update amplitude of each dimensional parameter during the model training cycle based on the feature extraction sequence, and adjusts the learning frequency parameter according to the weight and learning convergence of each feature parameter in the model. It combines the data distribution and inference delay to update the inference path of the model to obtain a state recognition model.

[0011] As a further solution of the present invention, the key update timing table includes a rotation period value, a key update time point, and a synchronization verification tag; the denoised signal data specifically includes a frequency segmentation structure, an energy suppression identifier, and a filtering processing result; the feature extraction sequence includes a sampling interval parameter, a device state feature group, and a change rate tag; and the state recognition model specifically refers to parameter convergence information, an inference path structure, and a feature dimension weight table.

[0012] As a further solution of the present invention, the data verification module includes:

[0013] The data fluctuation extraction submodule obtains the monitoring records of sensor nodes, extracts the sensor data sequence change amplitude, sampling interval fluctuation value, and continuous change rate of each sensor node, analyzes the fluctuation trend and amplitude of the data sequence, and generates data fluctuation characteristic indicators;

[0014] The node operation frequency analysis submodule analyzes the node operation frequency and data synchronization offset according to the data fluctuation characteristic index, calculates the node activity frequency by evaluating the node load and working time, and generates node operation frequency data;

[0015] The key update period adjustment submodule adjusts the key rotation period based on the node operation frequency data and the data synchronization offset value, and generates a key update time sequence table.

[0016] As a further solution of the present invention, the signal noise reduction module includes:

[0017] The data frequency distribution acquisition submodule calls the key update time sequence table, collects the data frequency distribution value and the corresponding signal energy value in the sliding window, extracts the frequency density, peak frequency band position, and signal energy mean of each data window, and establishes the frequency energy distribution interval;

[0018] The energy value trend analysis submodule analyzes the time series of the signal energy mean of each frequency segment according to the frequency energy distribution interval, and calculates the signal frequency bandwidth change trend and high-frequency energy growth rate in combination with the movement distance and direction identification of the frequency concentration position in adjacent time periods to obtain the frequency domain change trend parameters;

[0019] The signal denoising execution submodule calls the frequency domain change trend parameter, adjusts the frequency band segmentation range and filtering threshold in signal processing according to the signal energy change rate and the frequency concentration segment displacement, performs denoising filtering on each frequency band signal and reconstructs the signal output to generate denoised signal data.

[0020] As a further solution of the present invention, the feature extraction module includes:

[0021] The channel fluctuation acquisition submodule calls the denoised signal data to obtain the change amplitude, update frequency, and fluctuation trend difference value of each sensor data channel in the sliding window. Combined with the data change direction difference and frequency change of each channel in continuous time slices, it constructs a fluctuation identification label sequence for each channel and generates channel fluctuation statistics.

[0022] The dimensional pressure construction submodule extracts the change amplitude, offset interval, and intensive fluctuation frequency of each characteristic dimension in the window overlap area based on the channel fluctuation statistics, uses the change rate normalization value and the offset superposition value to extract the change pressure value, and divides the characteristic dimension groups according to the numerical interval to obtain the characteristic dimension pressure interval value;

[0023] The feature structure output submodule calls the pressure interval value of the feature dimension, calculates the sampling interval adjustment value according to the feature dimension of each pressure group and the corresponding data change rate, adjusts the sampling frequency parameter of each dimension, extracts the data features of the operating parameters of various IoT devices, and generates a feature extraction sequence.

[0024] As a further solution of the present invention, the model building module includes:

[0025] The feature change extraction submodule extracts the data change range and parameter update amplitude of each feature dimension within the model training cycle based on the feature extraction sequence, calculates the change frequency and value range span of each dimension within the time axis, compares the change amplitude and frequency difference, calculates the feature importance score, and obtains the feature change weight coefficient;

[0026] The learning frequency control submodule calls the feature change weight coefficient, assigns weights according to the importance score of each dimension, adjusts the learning frequency parameters based on the learning convergence situation, and obtains a feature learning frequency parameter group;

[0027] The inference path update submodule learns the frequency parameter group according to the feature, updates the inference path of the model by evaluating the data distribution and inference delay during the training process, identifies the state of the IoT device in real time, and builds a state recognition model.

[0028] As a further solution of the present invention, the specific formula for calculating the importance score of the feature is:

[0029]

[0030] Calculate feature importance scores;

[0031] Among them, I b represents the importance score of feature dimension b, Δx b，a is the value change of feature dimension b in the ath time slice, R bis the value range of feature dimension b during the training cycle, is the average value range of all feature dimensions, A is the total number of observation time slices of feature dimension b, A is the total number of time slices in the model training cycle, a is the time slice number of the current sampling, and b is the feature dimension number.

[0032] As a further embodiment of the present invention, the system further comprises:

[0033] The resource scheduling module obtains the state recognition model, analyzes the data input required for multiple reasoning tasks, the model scale, and the node load status, analyzes the task processing time, load pressure, and processing response order, adjusts the task queue and resource allocation according to the task urgency and node processing capacity, and generates a resource allocation scheduling result;

[0034] The resource allocation and scheduling results include a task queue sequence table, a node resource mapping table, and a response delay distribution diagram.

[0035] As a further solution of the present invention, the resource scheduling module includes:

[0036] The task input evaluation submodule calls the state recognition model to obtain the data input amount and the number of model structure layers corresponding to each reasoning task, performs normalization conversion on the data processing amount and structure depth of each task, constructs a task scale parameter set based on the conversion value, and generates an inference task scale parameter group;

[0037] The node load judgment submodule extracts the operation frequency, data queue length, and resource utilization rate of each computing node in the current cycle based on the inference task scale parameter group, and calculates the load pressure index value of each node by combining the number of tasks to be processed and the interval between node task completions to obtain the node operation load coefficient;

[0038] The resource allocation update submodule calls the node operation load coefficient, combines the task urgency level, performs matching calculation on the task priority sequence and the node processing capacity, adjusts the task queue position according to the task urgency level and the offset of the node pressure value, and adjusts the computing resource allocation to generate the resource allocation scheduling result.

[0039] As a further solution of the present invention, the specific formula for calculating the load pressure index value of each node is:

[0040]

[0041] Calculate the node load pressure index value;

[0042] Among them, P m represents the operating load factor of the mth computing node, f mRepresents the frequency of task execution of the node in the current cycle, F m Represents the maximum operating frequency threshold of a single node set by the system, q m Represents the number of tasks currently queued for processing on the node, Q m Represents the task quantity threshold, u m Represents the current node resource occupancy rate, Represents the arithmetic mean of resource utilization of all nodes in the same period, U m Represents the reference upper limit of node resource utilization, t m Represents the actual completion interval of the node's most recent task, Δt m，q Represents the change in the completion time of the p tasks before this node, T m represents the maximum allowed time interval for the node to process tasks, p is the number of participants in the node's historical task completion interval, q is the index number of the queued task on the current node, and m is the number of the computing node in the current IoT platform.

[0043] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0044] The key rotation rhythm is dynamically adjusted through data fluctuation characteristics and node synchronization offset to enhance the integrity and security of data transmission. The frequency distribution and energy trend are linked to set the filtering parameters to improve the denoising adaptability of non-stationary signals. The sampling frequency is adjusted according to the rate of change, which improves the extraction accuracy of multi-source device operation characteristics. The training frequency is dynamically allocated based on the feature weight and learning convergence status, which optimizes the distribution of training resources and convergence efficiency, enhances the responsiveness of the model structure to task pressure, matches the scheduling order according to the task urgency and the node pressure bearing capacity, and improves the coordination and real-time performance of resource utilization in a concurrent task environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 is a system flow chart of the present invention;

[0047] Figure 2 Schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0049] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0050] In the embodiments of the present invention, the terms "image" and "picture" may be used interchangeably. It should be noted that, when the distinction between them is not emphasized, their intended meanings are the same. The terms "of," "corresponding," and "corresponding" may be used interchangeably. It should be noted that, when the distinction between them is not emphasized, their intended meanings are the same.

[0051] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0052] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0053] The embodiment of the present invention provides an AI algorithm engine system applied to the Internet of Things platform, please refer to Figures 1 to 2 The present invention provides a technical solution, wherein the AI algorithm engine system applied to the Internet of Things platform includes:

[0054] The data verification module obtains the monitoring records of the sensor nodes, performs statistical analysis on the data fluctuations in each time period, extracts the data fluctuation characteristic indicators, and adjusts the key rotation cycle based on the operation frequency of the transmission node and the data synchronization offset value to generate a key update sequence list;

[0055] The signal denoising module calls the key update sequence table to collect the data frequency distribution and corresponding signal energy values within the sliding window. Based on the changing trend of the high-frequency energy value and the moving direction of the frequency concentration area, it adjusts the frequency segmentation range and filter threshold settings in the signal processing process to obtain denoised signal data.

[0056] The feature extraction module uses denoised signal data and the data change rate to adjust the data sampling frequency parameters, extracting the data features of various IoT device operating parameters and generating a feature extraction sequence;

[0057] The model building module extracts the variation range and update amplitude of each dimension parameter during the model training cycle based on the feature extraction sequence. It also adjusts the learning frequency parameter based on the weight of each feature parameter in the model and the learning convergence status. It updates the inference path of the model by combining the data distribution and inference delay to obtain a state recognition model.

[0058] The resource scheduling module obtains the state recognition model, analyzes the data input volume, model scale, and node load status required for multiple inference tasks, analyzes the task processing time, load pressure level, and processing response order, and adjusts the task queue and resource allocation according to the task urgency and node processing capacity to generate resource allocation scheduling results.

[0059] The key update sequence table includes the rotation period value, key update time point, and synchronization verification tag. The denoised signal data specifically includes the frequency segmentation structure, energy suppression flag, and filtering processing results. The feature extraction sequence includes the sampling interval parameters, device status feature group, and change rate tag. The state recognition model specifically refers to parameter convergence information, inference path structure, and feature dimension weight table. The resource allocation scheduling results include the task queue sequence table, node resource mapping table, and response delay distribution diagram.

[0060] The data verification module includes:

[0061] The data fluctuation extraction submodule obtains the monitoring records of sensor nodes, extracts the sensor data sequence change amplitude, sampling interval fluctuation value, and continuous change rate of each sensor node, analyzes the fluctuation trend and amplitude of the data sequence, and generates data fluctuation characteristic indicators;

[0062] The data fluctuation extraction submodule calls the data log file uploaded by the sensor node, reorders the monitoring data of each node according to the timestamp order, and uniformly resamples the data at an interval of 10ms. It extracts the absolute value of the difference between the values of each two adjacent sampling points, calculates the change amplitude of the data sequence, and calculates the maximum, minimum and average values of the change amplitude. It then extracts the timestamp difference between each two adjacent data points, calculates the standard deviation of the time interval as the sampling interval fluctuation value, and further uses the sliding window method to extract the data change rate to obtain the local change rate mean. The formula is:

[0063] V i =|X i -X i-1 |;

[0064] Calculate the data change range, where V i is the variation range of the i-th data point, X i is the value of the i-th data point, X i-1 is the value of the i-1th data point. Set X i-1 =22,X i=25, substitute the set value into the calculation:

[0065] V i =|25-22|=3;

[0066] The calculation results show that the change amplitude of the current data point relative to the previous data point is 3. Then this method is used to calculate the change amplitude of the entire sequence data, and these amplitude values are statistically analyzed to extract the fluctuation trend and amplitude characteristics after standardization to form a data fluctuation characteristic index.

[0067] The node operation frequency analysis submodule analyzes the node operation frequency and data synchronization offset based on the data fluctuation characteristic indicators. By evaluating the node load and working hours, it calculates the node activity frequency and generates node operation frequency data.

[0068] The node operation frequency analysis submodule counts the total number of data submitted and the number of high-volatility data within the node monitoring period based on the data fluctuation characteristic index, calculates the operation time based on the difference between the first recording time and the last recording time of the node, calculates the node activity based on the operation time, and then calculates the proportion of the number of high-volatility data to the total number of submissions to obtain the node operation proportion. Finally, the mean value of the difference between the node data timestamp and the system standard time is extracted as the synchronization offset value. The node activity, operation proportion, and synchronization offset value are used to construct a node activity frequency evaluation system. The formula is used:

[0069]

[0070] Compute the running ratio of nodes, where R n is the running proportion of node n, M n The number of high-volatility data submitted for node n, N n The total number of data submitted for node n. Set M n =80, N n =100, substitute the set value for calculation:

[0071]

[0072] The calculation results show that the number of times node n submits high-volatility data accounts for 80% of the total data submission times. This operation proportion is then normalized and combined with the node activity and synchronization offset value through weight distribution to form the node operation frequency data.

[0073] The key update cycle adjustment submodule adjusts the key rotation cycle based on the node operation frequency data and the data synchronization offset value, and generates a key update sequence table;

[0074] The key update cycle adjustment submodule extracts the activity frequency value and data synchronization offset mean of each node based on the node operation frequency data, and combines these two indicators to evaluate the node risk level. Nodes with an activity frequency greater than 0.8 and an offset mean greater than 100ms are determined to be high risk, with a cycle set to 600 seconds. Nodes with an activity frequency between 0.5 and 0.8 and an offset less than 100ms are medium risk, with a cycle set to 1200 seconds. Nodes with an activity frequency less than 0.5 or an offset less than 50ms are low risk, with a cycle set to 1800 seconds. A rotation cycle mapping table is formulated accordingly. If the risk level fluctuates by more than one level in three consecutive risk level assessments, a key rotation is performed in advance, with an advance margin of 20% of the current cycle. The formula is used:

[0075] F n =0.6×H n +0.4×P n ;

[0076] Calculate the node risk judgment index, where F n is the risk index value of node n, H n is the node activity frequency value, P n The node offset mean normalized value (based on 500ms). Set H n =0.85, the mean offset is 120ms, then after normalization P n =120 / 500=0.24, substitute the set value for calculation:

[0077] F n =0.6×0.85+0.4×0.24=0.51+0.096=0.606;

[0078] The calculation results show that the node risk index value is 0.606. A value exceeding 0.6 corresponds to a high risk level. The cycle is set to 600 seconds. The risk levels of all nodes are calculated in turn to form a key update sequence list.

[0079] The signal noise reduction module includes:

[0080] The data frequency distribution acquisition submodule calls the key update sequence table to collect the data frequency distribution value and corresponding signal energy value in the sliding window, extracts the frequency density, peak frequency band position, and signal energy mean of each data window, and establishes the frequency energy distribution range;

[0081] The data frequency distribution acquisition submodule calls the key update sequence list, collects the signal data segments in each sliding window according to the set time window width, and performs frequency domain transformation operations to extract the data frequency distribution characteristics. First, the number of occurrences of each frequency component is obtained, the frequency density value is calculated and normalized, and the position of the main peak in the frequency density is used as the frequency peak segment. By extracting the square accumulation of the amplitude corresponding to each frequency point in the frequency interval, the total signal energy is calculated and the average value is taken as the signal energy mean. The frequency density, the frequency main peak segment position and the signal energy mean all constitute the frequency domain data characteristics of the current window. In different frequency intervals, the continuous frequency range is delineated and the average energy value is extracted to form a frequency energy distribution sequence, and the energy distribution results in adjacent windows are integrated in the time dimension to establish the frequency energy distribution interval. The sampling window length is set to 200ms, the frequency resolution is 5Hz, the upper limit of the frequency interval division is 100Hz, and it is divided into 20 frequency segments. The energy mean of each segment is calculated using the amplitude square and is expressed by the following formula:

[0082]

[0083] Among them, E k is the average energy value of the kth frequency segment, N k is the number of sampling points in the frequency band, A k，j is the amplitude of the jth sampling point. Set N k =4, A k，1 =3, A k，2 =5, A k，3 =4, A k，4 =2, substitute the set value to calculate:

[0084]

[0085] The calculation results show that the average energy value of the frequency segment is 13.5. This value is used as the energy level corresponding to the frequency interval, and is further linked with the mean values of other frequency segments to divide the frequency energy distribution interval.

[0086] The energy value trend analysis submodule analyzes the time series of the signal energy mean of each frequency segment according to the frequency energy distribution interval. Combined with the movement distance and direction identification of the frequency concentration position in adjacent time periods, it calculates the signal frequency bandwidth change trend and high-frequency energy growth rate to obtain the frequency domain change trend parameters.

[0087] The energy value trend analysis submodule obtains the average energy value sequence of each frequency segment in different time windows based on the frequency energy distribution interval, extracts the energy value in chronological order and constructs a time series, calculates the energy value change difference of each frequency segment in adjacent time windows, counts the energy change rate of the high-frequency part within the frequency band bandwidth, and marks its offset direction based on the lateral displacement value within the window of the main peak position of the frequency segment. The calculation of the high-frequency energy growth rate is achieved by taking the change amplitude of the high-frequency energy mean value in two consecutive time periods and dividing it by the energy value of the previous period. The frequency main peak offset is taken as the main peak position difference and the direction is assigned. The high-frequency energy growth rate uses the following formula:

[0088]

[0089] Among them, G h is the energy growth rate in the high frequency band, E t1 is the energy mean of the high frequency band in the previous time window, E t2 is the energy average of the high frequency band in the current time window. Set E t1 =10.0,E t2 =14.0, substitute the set value into the calculation:

[0090]

[0091] The calculation results show that the growth rate of high-frequency band energy in the current window is 0.4, which means that the energy of the high-frequency band has a continuous strengthening trend. Combined with the movement amount and direction of the main frequency peak in the window (for example, the main peak of the previous window is at 35Hz, and the main peak of this window is at 40Hz, then the offset is +5Hz), the frequency domain change trend parameter is formed. Finally, the energy change rate and main frequency offset direction of each frequency band in multiple time windows are combined to form a frequency domain change trend parameter set.

[0092] The signal denoising execution submodule calls the frequency domain change trend parameter, adjusts the frequency band segmentation range and filtering threshold in signal processing according to the signal energy change rate and the frequency concentration segment displacement, performs denoising filtering on each frequency band signal, and reconstructs the signal output to generate denoised signal data;

[0093] The signal denoising execution submodule calls the frequency domain change trend parameter, marks the frequency segment with an energy growth rate greater than 0.3 in the frequency domain as the noise-dominated segment, and the frequency segment with a main peak offset greater than 10Hz as the unstable frequency band segment. The above two parameters are combined to evaluate the signal stability of each frequency segment, tighten the frequency division range to 3Hz for unstable frequency bands, and expand the frequency coverage to 7Hz for stable frequency bands. The filtering threshold is compared with the energy average value of each frequency band and the energy reference value set by the system. If the energy value of a frequency band is more than 30% higher than the reference value, the filtering intensity of the segment is set to strong filtering. After the frequency band is divided and the filtering threshold is set, the filtering operation is performed on the signal of each frequency band to reconstruct the overall time domain signal. The filtering intensity parameter is calculated based on the difference between the energy value and the energy reference value, using the formula:

[0094]

[0095] Among them, S f is the filter strength parameter, E k is the average energy value of a certain frequency segment, E b is the energy reference value. Set E k =26, E b =20, substitute the set value to calculate:

[0096]

[0097] The calculation results show that the current frequency band has a 30% increase in energy compared to the baseline. A high-intensity filter setting is used, and a tightening threshold is applied to the frequency band to complete the time domain signal reconstruction, and the final output is the denoised signal data.

[0098] The feature extraction module includes:

[0099] The channel fluctuation acquisition submodule calls the denoised signal data to obtain the change amplitude, update frequency, and fluctuation trend difference value of each sensor data channel in the sliding window. Combining the data change direction difference and frequency change of each channel in continuous time slices, it constructs a fluctuation identification label sequence for each channel and generates channel fluctuation statistics.

[0100] The channel fluctuation acquisition submodule calls the denoised signal data to obtain the variation amplitude, update frequency and fluctuation trend difference of each sensor data channel in the sliding window. First, the maximum and minimum values in the continuous time slices are collected from each channel and the difference operation is performed to obtain the variation amplitude V. The number of refreshes of the data record per unit time is counted as the update frequency F. The number of data direction changes between each pair of adjacent time slices is calculated to obtain the direction change frequency D. Then, the weight values α, β, and γ are set by the normalization coefficient to construct the channel fluctuation intensity value using the formula:

[0101] W = α·V + β·F + γ·D;

[0102] Among them, W is the channel fluctuation intensity value, V is the channel change amplitude, F is the data update frequency, D is the direction change frequency, α is the amplitude weight coefficient, β is the update frequency weight coefficient, and γ is the direction difference weight coefficient.

[0103] Setting: α=0.5, β=0.3, γ=0.2, V=3.4, F=5, D=4, substitute into the calculation:

[0104] W=0.5·3.4+0.3·5+0.2·4=1.7+1.5+0.8=4.0;

[0105] The fluctuation intensity value is 4.0. Combined with the fluctuation sequence within the time slice, if W>3.5 in three consecutive slices, the channel is assigned a high-frequency fluctuation label, otherwise it is assigned a stable or neutral label, and finally the channel fluctuation statistics are generated.

[0106] The dimensional pressure construction submodule extracts the change amplitude, offset interval, and intensive fluctuation frequency of each feature dimension in the window overlap area based on the channel fluctuation statistics. It uses the change rate normalization value and the offset superposition value to extract the change pressure value, and divides the feature dimension groups according to the numerical interval to obtain the feature dimension pressure interval value.

[0107] The dimensional pressure construction submodule collects the change amplitude A, offset interval B, and intensive fluctuation frequency C of each characteristic dimension in the overlapping area of the sliding window based on the channel fluctuation statistics. Each parameter is first normalized to the maximum value and then directly superimposed to obtain the change pressure value P using the formula:

[0108]

[0109] Among them, P is the pressure value, A is the amplitude of change, B is the offset, C is the frequency of intensive fluctuations, and A max 、B max 、C max is the maximum value of each parameter in the current window.

[0110] Setting: A=1.5, B=1.0, C=3, A max =2.0,B max =1.5,C max =5, substitute into the calculation:

[0111]

[0112] The calculated pressure change value is 2.02. According to the pressure interval rule, [0,1] is the low pressure group, (1,2] is the medium pressure group, and (2,+∞) is the high pressure group. Therefore, this feature dimension is divided into the high pressure group to obtain the pressure interval value of the feature dimension.

[0113] The feature structure output submodule calls the characteristic dimension pressure interval value, calculates the sampling interval adjustment value based on the characteristic dimension of each pressure group and the corresponding data change rate, adjusts the sampling frequency parameter of each dimension, extracts the data features of various IoT device operating parameters, and generates a feature extraction sequence;

[0114] The feature structure output submodule calls the characteristic dimension pressure interval value. According to the pressure grouping result, the sampling interval of the dimension parameters in each group is adjusted in combination with its data change rate R. The weighted inverse relationship between the sampling adjustment factor S and the change pressure P is set as follows:

[0115]

[0116] Among them, S is the sampling adjustment factor, R is the rate of change, and P is the pressure value.

[0117] Setting: R = 1.2, P = 2.02, substitute into the calculation:

[0118]

[0119] The sampling adjustment factor is 0.802. If the basic sampling interval is 1 second, the sampling period of this dimension will be updated to 0.802 seconds. All dimensions will be sampled according to the updated sampling interval and then reconnected and rearranged to extract the data features of various IoT device operating parameters and generate a feature extraction sequence.

[0120] Model building modules include:

[0121] The feature change extraction submodule extracts the data change range and parameter update amplitude of each feature dimension within the model training cycle based on the feature extraction sequence, calculates the change frequency and value range span of each dimension within the time axis, compares the change amplitude and frequency difference, calculates the feature importance score, and obtains the feature change weight coefficient;

[0122] The specific formula for calculating the importance score of a feature is:

[0123]

[0124] Calculate feature importance scores;

[0125] Among them, I b represents the importance score of feature dimension b, Δx b，a is the value change of feature dimension b in the ath time slice, R b is the value range of feature dimension b during the training cycle, is the average value range of all feature dimensions, A is the total number of observation time slices of feature dimension b, A is the total number of time slices in the model training cycle, a is the time slice number of the current sampling, and b is the feature dimension number.

[0126] formula:

[0127]

[0128] Detailed explanation of the formula and the process of formula calculation and derivation:

[0129] The formula is used to calculate the importance score of each feature dimension. The result is used to characterize the intensity and relative fluctuation of the feature during the model training cycle, and serves as a quantitative basis for weight adjustment in the feature selection and training process.

[0130] Parameter meaning and setting value:

[0131] A is the total number of observation time slices of feature dimension b, which is set to 5 and reflects the number of time slice sampling points of a certain feature under the sliding window;

[0132] Δx b，a is the change value of the feature in the ath time slice compared with the previous time slice. The collected value sequence is 0.4, 0.3, 0.6, 0.2, 0.5, and the corresponding unit is the percentage of the sensor's standard range;

[0133] R b Set R as the value range of the feature dimension during the training cycle. b =3.0;

[0134] For the average value range of all feature dimensions, set

[0135] Substitute the parameters into the formula for calculation:

[0136]

[0137] The result of 0.1374 indicates that this feature dimension has a medium-frequency value change within the sliding period. At the same time, compared with all feature dimensions, its change amplitude is closer to the mean. The value is used as the feature change weight coefficient in the subsequent feature sorting and weight adjustment process.

[0138] The learning frequency control submodule calls the feature change weight coefficient, assigns weights according to the importance score of each dimension, adjusts the learning frequency parameters based on the learning convergence situation, and obtains the feature learning frequency parameter group;

[0139] The learning frequency control submodule calls the feature change weight coefficient, assigns the training frequency priority to each feature dimension according to the weight score, and further evaluates its learning adequacy based on the convergence state, obtains the training error change rate and iterative stability index in the current cycle, and judges whether the learning frequency of each feature needs to be increased or decreased based on this index. The training error change rate is obtained by calculating the loss difference of each batch and the number of training steps, and the stability factor is calculated by calculating the gradient variance change coefficient within 5 consecutive batches. The learning frequency parameter value uses the formula:

[0140]

[0141] Among them, F L is the learning frequency parameter value, W is the feature weight, E is the convergence error change rate, and S is the stability factor; set W = 0.75, E = 0.08, S = 0.2, and substitute the set values for calculation:

[0142]

[0143] The calculation results show the learning frequency rhythm that the feature should adopt in the current training process, thereby forming a feature learning frequency parameter group.

[0144] The inference path update submodule updates the model's inference path based on the feature learning frequency parameter group and evaluates the data distribution and inference latency during the training process, thereby identifying the state of IoT devices in real time and building a state recognition model.

[0145] The inference path update submodule traverses the feature set corresponding to each path in the model based on the feature learning frequency parameter group, and records the number of nodes and average response time experienced by each path during inference. It combines and evaluates the skewness coefficient of each feature distribution in the training data with the node response time in the current inference path, and calculates the inference load ratio of each path. The path response delay is averaged by the total node response time, and the skewness coefficient is obtained by the skewness of the frequency distribution of each feature in the training set. The inference load value is calculated using the formula:

[0146]

[0147] Where P is the path load value, D is the inference delay, K is the skewness coefficient, and F L For learning frequency; set D = 320ms, K = 1.1, F L =0.05, substitute the set value into the calculation:

[0148]

[0149] The calculation results show the response burden degree corresponding to each path. Based on the results, the paths are sorted and optimized to complete the construction of the state recognition model.

[0150] The resource scheduling module includes:

[0151] The task input evaluation submodule calls the state recognition model to obtain the data input volume and model structure layer number corresponding to each reasoning task, performs normalization conversion on the data processing volume and structure depth of each task, constructs a task scale parameter set based on the conversion values, and generates an inference task scale parameter group;

[0152] The task input evaluation submodule calls the state recognition model, extracts the data input volume and model structure layer number corresponding to each inference task, calculates the data processing volume and structure depth of each task, and uniformly converts them into task scale parameters for task load evaluation. In actual operation, the input data length and data type dimension are first extracted from the task call request and the number of valid hierarchical structures of the model associated with the task is recorded. Then, the two contents are normalized by proportional mapping. For example, if the task input volume is D and the number of model layers is L, the maximum input volume D is set. max 2000 units, maximum number of layers L max If the number of layers is 30, the unified conversion adopts the following formula:

[0153]

[0154] Among them, S is the task scale value, D is the task input data volume, and D max is the maximum standard value of the input data, L is the number of model layers, L max It is the standard upper limit of the number of model layers.

[0155] Assume D = 1500, L = 20, and substitute the calculation into:

[0156]

[0157] The single task scale value is 1.4167. Then, the scale values corresponding to each task are recorded in sequence, and a task scale parameter set is constructed to obtain the inference task scale parameter group.

[0158] The node load judgment submodule extracts the operation frequency, data queue length, and resource utilization rate of each computing node in the current cycle based on the inference task scale parameter group. It then calculates the load pressure index value of each node by combining the number of tasks to be processed and the interval between node task completions to obtain the node operation load coefficient.

[0159] The specific formula for calculating the load pressure index value of each node is:

[0160]

[0161] Calculate the node load pressure index value;

[0162] Among them, P m represents the operating load factor of the mth computing node, f m Represents the frequency of task execution of the node in the current cycle, F m Represents the maximum operating frequency threshold of a single node set by the system, q m Represents the number of tasks currently queued for processing on the node, Q m Represents the task quantity threshold, u m Represents the current node resource occupancy rate, Represents the arithmetic mean of resource utilization of all nodes in the same period, U m Represents the reference upper limit of node resource utilization, t m Represents the actual completion interval of the node's most recent task, Δt m，q Represents the change in the completion time of the p tasks before this node, T m represents the maximum allowed time interval for the node to process tasks, p is the number of participants in the node's historical task completion interval, q is the index number of the queued task on the current node, and m is the number of the computing node in the current IoT platform.

[0163] formula:

[0164]

[0165] Detailed explanation of the formula and the process of formula calculation and derivation:

[0166] The formula is used to calculate the operating load factor of a single computing node in the IoT platform. The result is used to measure the task processing pressure and resource usage fluctuation of the current node, which serves as the basis for subsequent task scheduling and resource reallocation.

[0167] Parameter meaning and setting value:

[0168] f m is the actual running frequency of node m in the current cycle, which is set to 180 times / cycle;

[0169] F m The maximum operation frequency threshold for this type of node is set to 250 times / cycle;

[0170] q m The length of the task queue of the current node is set to 8;

[0171] Q m is the maximum queued task threshold, set to 12;

[0172] u m The CPU resource usage of the current node is set to 0.84;

[0173] is the average CPU usage of all nodes in the same period, set to 0.67;

[0174] U m The upper limit of CPU usage of the reference node is set to 0.95;

[0175] t m The time interval for the current node to complete the previous task is set to 2.8 seconds;

[0176] Δt m，q is the change in the completion time of the node's nearly p tasks, set to 0.2, 0.3, and 0.1 seconds;

[0177] T m The maximum allowed interval for task processing for this node is set to 5 seconds;

[0178] p is the number of samples of the task completion interval in the recent stage, which is set to 3.

[0179] Substitute the parameters into the formula for calculation:

[0180]

[0181] P m =0.41965+0.3631≈0.78275;

[0182] The result of 0.78275 indicates that the current node's operating load is in the medium-to-high range. Compared with the reference full load standard of 1, this value indicates that the current node's resource usage is close to the system load control threshold. Subsequent scheduling should consider reducing task allocation to this node. The calculated operating load coefficient is the main basis for adjusting the priority of subsequent task allocation.

[0183] The resource allocation update submodule calls the node operation load coefficient, combines the task urgency level, performs matching calculations on the task priority order and node processing capacity, adjusts the task queue position and computing resource allocation according to the offset between the task urgency level and the node pressure value, and generates resource allocation scheduling results;

[0184] The resource allocation update submodule calls the node operation load coefficient, and combines the task urgency level to calculate the task priority queue position adjustment value and the corresponding computing resource reallocation ratio. During the processing, the task priority level r ranges from 1 to 5, representing five levels from low to high. The node load coefficient P corresponds to the one just calculated. Let the total resource be R total , the computing resource weight is allocated according to the following formula:

[0185]

[0186] Among them, R i The resource value assigned to the current task, Rtotal is the total resources that can be allocated to the node, r is the task urgency level, and P is the current load factor of the node.

[0187] Set R total = 100 resource units, the task level is r = 4, and the node pressure is P = 2.5. Substituting into the equations, we get:

[0188]

[0189] This indicates that the resource allocation for the task on this node requires 160 units to be allocated from the total. After the corresponding scheduling queue is reordered, the task queue position order and node resource configuration are updated to generate the resource allocation scheduling result.

[0190] If you need to continue expanding the content of other modules, please continue to send the corresponding paragraphs or modules.

[0191] The above embodiments can be implemented in whole or in part through software, hardware (such as circuits), firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the processes or functions according to the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired method (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, or magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0192] It should be understood that the term "and / or" as used herein simply describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. A and B can be singular or plural. Furthermore, the character " / " as used herein generally indicates an "or" relationship between the associated objects, but it may also indicate an "and / or" relationship. For specific understanding, please refer to the context.

[0193] In this disclosure, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, "at least one of a, b, or c" can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or plural.

[0194] It should be understood that in various embodiments of the present invention, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0195] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0196] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0197] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interface, indirect coupling or communication connection of the device or unit, which can be electrical, mechanical or other forms.

[0198] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0199] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0200] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0201] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. The AI algorithm engine system applied to the Internet of Things platform is characterized by: The system comprises: The data verification module obtains the monitoring records of the sensor nodes, performs statistical analysis on the data fluctuations in each time period, extracts the data fluctuation characteristic indicators, and adjusts the key rotation cycle based on the operation frequency of the transmission node and the data synchronization offset value to generate a key update sequence list; The signal noise reduction module calls the key update time sequence table, collects the data frequency distribution and corresponding signal energy value in the sliding window, adjusts the frequency segment range and filter threshold setting in the signal processing process according to the change trend of the high-frequency energy value and the moving direction of the frequency concentration area, and obtains the denoised signal data; The feature extraction module uses the denoised signal data and the data change rate to adjust the sampling frequency parameters of the data, extracts the data features of the operating parameters of various IoT devices, and generates a feature extraction sequence; The model construction module extracts the variation range and update amplitude of each dimensional parameter during the model training cycle based on the feature extraction sequence, and adjusts the learning frequency parameter according to the weight and learning convergence of each feature parameter in the model. It combines the data distribution and inference delay to update the inference path of the model to obtain a state recognition model.

2. The AI algorithm engine system applied to the Internet of Things platform according to claim 1, characterized in that: The key update sequence table includes a rotation period value, a key update time point, and a synchronization verification tag; the denoised signal data specifically includes a frequency segmentation structure, an energy suppression identifier, and a filtering processing result; the feature extraction sequence includes a sampling interval parameter, a device state feature group, and a change rate tag; the state recognition model specifically refers to parameter convergence information, an inference path structure, and a feature dimension weight table.

3. The AI algorithm engine system applied to the Internet of Things platform according to claim 1, characterized in that: The data verification module includes: The data fluctuation extraction submodule obtains the monitoring records of sensor nodes, extracts the sensor data sequence change amplitude, sampling interval fluctuation value, and continuous change rate of each sensor node, analyzes the fluctuation trend and amplitude of the data sequence, and generates data fluctuation characteristic indicators; The node operation frequency analysis submodule analyzes the node operation frequency and data synchronization offset according to the data fluctuation characteristic index, calculates the node activity frequency by evaluating the node load and working time, and generates node operation frequency data; The key update period adjustment submodule adjusts the key rotation period based on the node operation frequency data and the data synchronization offset value, and generates a key update time sequence table.

4. The AI algorithm engine system applied to the Internet of Things platform according to claim 3 is characterized in that: The signal noise reduction module includes: The data frequency distribution acquisition submodule calls the key update time sequence table, collects the data frequency distribution value and the corresponding signal energy value in the sliding window, extracts the frequency density, peak frequency band position, and signal energy mean of each data window, and establishes the frequency energy distribution interval; The energy value trend analysis submodule analyzes the time series of the signal energy mean of each frequency segment according to the frequency energy distribution interval, and calculates the signal frequency bandwidth change trend and high-frequency energy growth rate in combination with the movement distance and direction identification of the frequency concentration position in adjacent time periods to obtain the frequency domain change trend parameters; The signal denoising execution submodule calls the frequency domain change trend parameter, adjusts the frequency band segmentation range and filtering threshold in signal processing according to the signal energy change rate and the frequency concentration segment displacement, performs denoising filtering on each frequency band signal and reconstructs the signal output to generate denoised signal data.

5. The AI algorithm engine system applied to the Internet of Things platform according to claim 4, characterized in that: The feature extraction module includes: The channel fluctuation acquisition submodule calls the denoised signal data to obtain the change amplitude, update frequency, and fluctuation trend difference value of each sensor data channel in the sliding window. Combined with the data change direction difference and frequency change of each channel in continuous time slices, it constructs a fluctuation identification label sequence for each channel and generates channel fluctuation statistics. The dimensional pressure construction submodule extracts the change amplitude, offset interval, and intensive fluctuation frequency of each characteristic dimension in the window overlap area based on the channel fluctuation statistics, uses the change rate normalization value and the offset superposition value to extract the change pressure value, and divides the characteristic dimension groups according to the numerical interval to obtain the characteristic dimension pressure interval value; The feature structure output submodule calls the pressure interval value of the feature dimension, calculates the sampling interval adjustment value according to the feature dimension of each pressure group and the corresponding data change rate, adjusts the sampling frequency parameter of each dimension, extracts the data features of the operating parameters of various IoT devices, and generates a feature extraction sequence.

6. The AI algorithm engine system applied to the Internet of Things platform according to claim 5, characterized in that: The model building module includes: The feature change extraction submodule extracts the data change range and parameter update amplitude of each feature dimension within the model training cycle based on the feature extraction sequence, calculates the change frequency and value range span of each dimension within the time axis, compares the change amplitude and frequency difference, calculates the feature importance score, and obtains the feature change weight coefficient; The learning frequency control submodule calls the feature change weight coefficient, assigns weights according to the importance score of each dimension, adjusts the learning frequency parameters based on the learning convergence situation, and obtains a feature learning frequency parameter group; The inference path update submodule learns the frequency parameter group according to the feature, updates the inference path of the model by evaluating the data distribution and inference delay during the training process, identifies the state of the IoT device in real time, and builds a state recognition model.

7. The AI algorithm engine system applied to the Internet of Things platform according to claim 6, characterized in that: The specific formula for calculating the importance score of the feature is: Calculate feature importance scores; Among them, I b represents the importance score of feature dimension b, Δx b，a is the value change of feature dimension b in the ath time slice, R b is the value range of feature dimension b during the training cycle, is the average value range of all feature dimensions, A is the total number of observation time slices of feature dimension b, A is the total number of time slices in the model training cycle, a is the time slice number of the current sampling, and b is the feature dimension number.

8. The AI algorithm engine system applied to the Internet of Things platform according to claim 1, characterized in that: The system further comprises: The resource scheduling module obtains the state recognition model, analyzes the data input required for multiple reasoning tasks, the model scale, and the node load status, analyzes the task processing time, load pressure, and processing response order, adjusts the task queue and resource allocation according to the task urgency and node processing capacity, and generates a resource allocation scheduling result; The resource allocation and scheduling results include a task queue sequence table, a node resource mapping table, and a response delay distribution diagram.

9. The AI algorithm engine system applied to the Internet of Things platform according to claim 8, characterized in that: The resource scheduling module includes: The task input evaluation submodule calls the state recognition model to obtain the data input amount and the number of model structure layers corresponding to each reasoning task, performs normalization conversion on the data processing amount and structure depth of each task, constructs a task scale parameter set based on the conversion value, and generates an inference task scale parameter group; The node load judgment submodule extracts the operation frequency, data queue length, and resource utilization rate of each computing node in the current cycle based on the inference task scale parameter group, and calculates the load pressure index value of each node by combining the number of tasks to be processed and the interval between node task completions to obtain the node operation load coefficient; The resource allocation update submodule calls the node operation load coefficient, combines the task urgency level, performs matching calculation on the task priority sequence and the node processing capacity, adjusts the task queue position according to the task urgency level and the offset of the node pressure value, and adjusts the computing resource allocation to generate the resource allocation scheduling result.

10. The AI algorithm engine system applied to the Internet of Things platform according to claim 9, characterized in that: The specific formula for calculating the load pressure index value of each node is: Calculate the node load pressure index value; Among them, P m represents the operating load factor of the mth computing node, f m Represents the frequency of task execution of the node in the current cycle, F m Represents the maximum operating frequency threshold of a single node set by the system, q m Represents the number of tasks currently queued for processing on the node, Q m Represents the task quantity threshold, u m Represents the current node resource occupancy rate, Represents the arithmetic mean of resource utilization of all nodes in the same period, U m Represents the reference upper limit of node resource utilization, t m Represents the actual completion interval of the node's most recent task, Δt m，q Represents the change in the completion time of the p tasks before this node, T m represents the maximum allowed time interval for the node to process tasks, p is the number of participants in the node's historical task completion interval, q is the index number of the queued task on the current node, and m is the number of the computing node in the current IoT platform.

Images