A grey model prediction system for building structure deformation analysis

Abstract

The application relates to the technical field of industrial information and data processing, and discloses a grey model prediction system for building structure deformation analysis, which comprises the following modules: a data preprocessing module for filtering industrial time series data sequences and storing the sequences in a first state register; an adaptive background value calculation module for determining an adaptive background value calibration weight according to a change gradient difference value, limiting the weight within 0.2-0.8 to generate a mean value background value sequence close to the weight; and a residual dynamic correction prediction module for constructing an observation matrix to solve a coefficient to output a trend prediction sequence, and calculating a deviation compensation value according to a prediction residual sequence to output deformation trend prediction data. The application can follow signal changes through adaptive weights, eliminate calculation lag and overshoot deviation in data mutation intervals, avoid divergence risks in parameter identification, and realize fast prediction response and deterministic threshold closed-loop arbitration on a low-power microprocessor.

Description

Technical Field

[0001] This invention belongs to the field of industrial information and data processing technology, and in particular relates to a gray model prediction system for building structure deformation analysis. Background Technology

[0002] Current data processing systems generally adopt the first-order univariate grey prediction method based on whitening differential equations to establish the temporal deformation trend evolution tensor. By performing first-order cumulative generation transformation on discrete historical shift sequences, the seemingly disordered discrete observation points are transformed into continuous fitting trajectories with monotonically exponential growth characteristics. This maintains low microprocessor computing power overhead under the constraints of low sample size and incomplete information, providing basic data flow support and trend fitting output.

[0003] However, when structural monitoring systems enter regions of strong non-stationary multivariate disturbances or face strong fluctuations due to multi-field coupling in alternating temperature fields, such classic data analysis mechanisms suffer from physical mismatches. Because they use a static weighting coefficient of 0.5 when generating background values ​​close to the mean, their implicit assumption of uniform and smooth data evolution becomes disconnected from the discrete physical waveforms of the data, which include sudden tremors and structural phase step abrupt changes. This causes the algorithm iteration flow within the central processing unit to face divergence problems in inverse matrix solving due to a sudden increase in local phase space curvature. Consequently, the trend envelope output by the parameter identification operator experiences continuous boundary overshoot, accumulating lag residuals and amplifying the conducted interference of random noise in time-series superposition calculations. This issue affects the synchronous sampling frequency of the front-end sensors and the back-end algorithm components. Besides the physical limitations of computing power load, existing grey model control methods have shortcomings in dealing with dynamic transient distortions. For example, Chinese invention patent CN111022124B discloses an advanced early warning method for short-term and long-term deformation of bridge and tunnel engineering. It improves the weight allocation of accumulated generated data by introducing optimization coefficients and corrects residuals using sinusoidal and periodic sequence functions. However, it is still limited by global empirical fitting of historical sequences. Since the optimization logic focuses on minimizing the sum of squares of global residuals, when faced with nonlinear step changes in building structures caused by transient external loads or environmental changes, the sensitivity of static weight adjustment is difficult to capture the microsecond-level time scale response. Moreover, the complex periodic function fitting causes a secondary expansion of microprocessor computing power overhead, resulting in phase lag in the prediction results.

[0004] Therefore, the technical problem to be solved by this invention is how to overcome the non-stationary time-series abrupt fitting lag caused by fixed background value coefficients in traditional grey prediction methods, and how to adaptively calculate weights by capturing the local fluctuation derivative gradient difference of the coupled variables of the target object and the environment, and providing boundary truncation gating hard constraints and dynamic residual feedback compensation closed loop for the whitening differential equation solution flow, so as to achieve high stability trend convergence with low computing power consumption. Summary of the Invention

[0005] This invention aims to solve the problems of fitting lag and solution divergence caused by non-stationary temporal abrupt changes in traditional gray prediction methods due to fixed background value coefficients.

[0006] In this technical solution, a grey model prediction system for building structure deformation analysis includes: The time-series data acquisition module is used to acquire industrial time-series data sequences that characterize the deformation state of building structures; The data standardization preprocessing module is connected to the time series data acquisition module through a channel. It is used to filter industrial time series data sequences and standardize and scale them to generate standard time series data sequences and store them in the first state register. The adaptive background value calculation module is connected to the data standardization preprocessing module through a channel. It is used to calculate the difference of change gradient based on the standard time series data sequence in the first state register to determine the adaptive background value calibration weight. The adaptive background value calibration weight is limited to 0.2 to 0.8. The weighted value is calculated based on the adaptive background value calibration weight and the standard time series data sequence, and a background value sequence with the nearest mean is generated in the background value register. The residual dynamic correction prediction module is connected to the adaptive background value calculation module through a channel. It is used to extract the nearest mean background value sequence from the background value register to construct the observation matrix. The development coefficient and driving coefficient of the differential equation are obtained by solving the least squares inverse matrix to output the trend prediction sequence. A synchronous difference sequence is established to capture the prediction residual sequence at the current sampling time. The second-order residual correction module calculates the deviation compensation value based on the prediction residual sequence and superimposes the deviation compensation value onto the output register to output the deformation trend prediction data.

[0007] Preferably, in the adaptive background value calculation module, the data boundary truncation module triggers a high-order truncation overwrite to 0.8 when the calculated adaptive background value calibration weight is greater than 0.8, and triggers a low-order truncation overwrite to 0.2 when the adaptive background value calibration weight is less than 0.2. The module also adds a deadlock prevention floating-point constant of 0.00001 preset in the status register of the adaptive background value calculation module to lock the fluctuation range of the adaptive background value calibration weight.

[0008] Preferably, the data standardization preprocessing module includes the following sub-steps: identifying and removing burst noise terms from the industrial time series data sequence using a time-domain median filtering algorithm; scaling the removed industrial time series data sequence using a max-min normalization algorithm so that the value range of the output standard time series data sequence is within the range of 0 to 1.

[0009] Preferably, the adaptive background value calculation module includes the following sub-steps: calculating the first-order difference quotient of adjacent time step data to obtain the deformation rate gradient; calculating the sliding window multiplication and addition value of the deformation rate gradient to generate a change gradient difference value reflecting the degree of temporal fluctuation.

[0010] Preferably, the residual dynamic correction prediction module includes the following sub-steps: establishing a residual tracking queue in system memory, reading the predicted output value and the actual deformation input feature value at the current sampling time in real time; comparing the predicted output value and the actual deformation input feature value, calculating the difference between the two as the predicted residual sequence and storing it in the residual tracking queue.

[0011] Preferably, the residual dynamic correction prediction module includes the following sub-steps: extracting the predicted residual sequence from the residual tracking queue as input parameters and sending it to the second-order residual correction module; fitting the development trend of the predicted residual sequence through the second-order residual correction module and outputting the deviation compensation value that changes with time.

[0012] Preferably, the system also includes a multi-level safety arbitration output module, which is connected to the residual dynamic correction prediction module via a data bus to receive deformation trend prediction data. The multi-level safety arbitration output module is used to establish a three-layer cascaded digital over-limit decision logic, comparing the predicted deformation amount in the deformation trend prediction data with fixed yellow warning threshold, orange warning threshold, and red safety limit threshold in the system memory. When the predicted deformation amount exceeds the yellow warning threshold for three consecutive control cycles, a first-level abnormal data is output; when it exceeds the orange warning threshold, a second-level deformation over-limit instruction is output; and when it exceeds the red safety limit threshold, a third-level system emergency stop control message is output.

[0013] Preferably, the residual dynamic correction prediction module includes the following sub-steps: extracting the sequence of adjacent mean background values ​​to form a column vector, and horizontally concatenating it with a one-dimensional constant vector of all 1s to construct an observation matrix; calculating the transpose product of the observation matrix to obtain a square matrix, and solving for the inverse matrix of the square matrix; using the inverse matrix, the transpose matrix of the observation matrix, and the first-order difference vector of the standard time series data sequence to calculate the multiplication mapping value, and identifying the development coefficient and driving coefficient of the differential equation.

[0014] Preferably, the system constrains the time-series flow through the system memory bus scheduling mechanism. A data bus feedback channel is configured between the deformation trend prediction data output by the residual dynamic correction prediction module and the gradient difference calculated by the adaptive background value calculation module. This allows the gradient difference calculated in the preceding steps to be directly used as the target input for the background value reorganization in the subsequent steps. Furthermore, the parameter identification result inversely constrains the input adaptive boundary of the second-order residual correction module, forming a memory storage architecture for a multi-dimensional feature time-series matrix.

[0015] Compared with existing technologies, the grey model prediction system for building structure deformation analysis of the present invention has the following advantages: 1. In the grey model prediction of building structure deformation analysis, a time-series sliding window data buffer with a time step length limited to 8 is set to simultaneously cache the time-series sequence of target deformation features and the time-series sequence of alternating temperature field features representing multi-field coupling interference factors in the environment. The data arbitration unit extracts the transient local gradient components of the first-order cumulative sequence of target deformation and the environmental feature sequence within adjacent time steps. The horizontal linear mapping logic is used to update the adaptive background value calibration weight in real time based on the nonlinear interaction characteristics of the gradients of the two, changing the design inertia of using constant background value weighting coefficients in traditional algorithms. This enables the system to adaptively follow the discrete distortion of the local first derivative of the signal when facing data jumps caused by sudden non-stationary disturbances, eliminating the calculation lag and overshoot deviation caused by fixed coefficients when the data sequence enters the phase change or nonlinear jump interval, and avoiding the risk of exponential divergence in the process of differential equation parameter identification.

[0016] 2. By introducing a hard data boundary truncation gate flow, the calculated adaptive background value calibration weights are set with a linear step-wise truncation judgment. When the weight value is greater than 0.8, a high-bit truncation action is triggered to forcibly overwrite and rewrite to 0.8. When the weight value is less than 0.2, a low-bit truncation action is triggered to forcibly overwrite and rewrite to 0.2. Combined with the pre-set anti-deadlock regularization floating-point constant 0.00001 in the status register, the fluctuation domain of the background value calibration weights is rigidly locked within a specific stable range. This provides convergence stability constraints for whitening differential equation parameter identification in the least squares inverse matrix solution process, avoiding computational deadlock caused by zero denominator or matrix singularity when the input data contains high-frequency random noise or sudden pulses. This improves the computational convergence performance of industrial computing nodes under the constraints of limited memory bandwidth and high concurrency throughput. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the gray model prediction system module for building structure deformation analysis of the present invention; Figure 2 This is a time-series data processing diagram of the grey model prediction system for building structure deformation analysis of this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0019] A grey model prediction system for building structure deformation analysis includes: The time-series data acquisition module is used to acquire industrial time-series data sequences that characterize the deformation state of building structures; The data standardization preprocessing module is connected to the time series data acquisition module through a channel. It is used to filter industrial time series data sequences and standardize and scale them to generate standard time series data sequences and store them in the first state register. The adaptive background value calculation module is connected to the data standardization preprocessing module through a channel. It is used to calculate the difference of change gradient based on the standard time series data sequence in the first state register to determine the adaptive background value calibration weight. The adaptive background value calibration weight is limited to 0.2 to 0.8. The weighted value is calculated based on the adaptive background value calibration weight and the standard time series data sequence, and a background value sequence with the nearest mean is generated in the background value register. The residual dynamic correction prediction module is connected to the adaptive background value calculation module through a channel. It is used to extract the nearest mean background value sequence from the background value register to construct the observation matrix. The development coefficient and driving coefficient of the differential equation are obtained by solving the least squares inverse matrix to output the trend prediction sequence. A synchronous difference sequence is established to capture the prediction residual sequence at the current sampling time. The second-order residual correction module calculates the deviation compensation value based on the prediction residual sequence and superimposes the deviation compensation value onto the output register to output the deformation trend prediction data.

[0020] Preferably, in the adaptive background value calculation module, the data boundary truncation module triggers a high-order truncation overwrite to 0.8 when the calculated adaptive background value calibration weight is greater than 0.8, and triggers a low-order truncation overwrite to 0.2 when the adaptive background value calibration weight is less than 0.2. The module also adds a deadlock prevention floating-point constant of 0.00001 preset in the status register of the adaptive background value calculation module to lock the fluctuation range of the adaptive background value calibration weight.

[0021] Preferably, the data standardization preprocessing module includes the following sub-steps: identifying and removing burst noise terms from the industrial time series data sequence using a time-domain median filtering algorithm; scaling the removed industrial time series data sequence using a max-min normalization algorithm so that the value range of the output standard time series data sequence is within the range of 0 to 1.

[0022] Preferably, the adaptive background value calculation module includes the following sub-steps: calculating the first-order difference quotient of adjacent time step data to obtain the deformation rate gradient; calculating the sliding window multiplication and addition value of the deformation rate gradient to generate a change gradient difference value reflecting the degree of temporal fluctuation.

[0023] Preferably, the residual dynamic correction prediction module includes the following sub-steps: establishing a residual tracking queue in system memory, reading the predicted output value and the actual deformation input feature value at the current sampling time in real time; comparing the predicted output value and the actual deformation input feature value, calculating the difference between the two as the predicted residual sequence and storing it in the residual tracking queue.

[0024] Preferably, the residual dynamic correction prediction module includes the following sub-steps: extracting the predicted residual sequence from the residual tracking queue as input parameters and sending it to the second-order residual correction module; fitting the development trend of the predicted residual sequence through the second-order residual correction module and outputting the deviation compensation value that changes with time.

[0025] Preferably, the system also includes a multi-level safety arbitration output module, which is connected to the residual dynamic correction prediction module via a data bus to receive deformation trend prediction data. The multi-level safety arbitration output module is used to establish a three-layer cascaded digital over-limit decision logic, comparing the predicted deformation amount in the deformation trend prediction data with fixed yellow warning threshold, orange warning threshold, and red safety limit threshold in the system memory. When the predicted deformation amount exceeds the yellow warning threshold for three consecutive control cycles, a first-level abnormal data is output; when it exceeds the orange warning threshold, a second-level deformation over-limit instruction is output; and when it exceeds the red safety limit threshold, a third-level system emergency stop control message is output.

[0026] Preferably, the residual dynamic correction prediction module includes the following sub-steps: extracting the sequence of adjacent mean background values ​​to form a column vector, and horizontally concatenating it with a one-dimensional constant vector of all 1s to construct an observation matrix; calculating the transpose product of the observation matrix to obtain a square matrix, and solving for the inverse matrix of the square matrix; using the inverse matrix, the transpose matrix of the observation matrix, and the first-order difference vector of the standard time series data sequence to calculate the multiplication mapping value, and identifying the development coefficient and driving coefficient of the differential equation.

[0027] Preferably, the system constrains the time-series flow through the system memory bus scheduling mechanism. A data bus feedback channel is configured between the deformation trend prediction data output by the residual dynamic correction prediction module and the gradient difference calculated by the adaptive background value calculation module. This allows the gradient difference calculated in the preceding steps to be directly used as the target input for the background value reorganization in the subsequent steps. Furthermore, the parameter identification result inversely constrains the input adaptive boundary of the second-order residual correction module, forming a memory storage architecture for a multi-dimensional feature time-series matrix.

[0028] Example 1: The current time-series data acquisition module acquires an industrial time-series data sequence characterizing the deformation state of the building structure. A data standardization preprocessing module connected to the time-series data acquisition module receives the industrial time-series data sequence and uses a time-domain median filtering algorithm to identify and remove burst noise terms from the industrial time-series data sequence. Simultaneously, a max-min normalization algorithm is used to scale the industrial time-series data sequence after removing burst noise terms, ensuring that the output standard time-series data sequence's value range is within the 0-1 range. The standard time-series data sequence is then stored in the first state register to reduce the transmission interference of high-frequency random pulse disturbances to the parameter identification operator. The adaptive background value calculation module, connected to the data standardization preprocessing module, extracts the standard time-series data sequence from the first state register and calculates the first-order difference quotient of adjacent time-step data to obtain the deformation rate gradient. Then, within a time-series sliding window data buffer with a time-step length limited to 8, a sliding window multiplication and addition value is calculated on the deformation rate gradient to generate a gradient difference reflecting the degree of time-series fluctuation. The data arbitration unit extracts the transient local gradient component of the first-order deformation accumulation sequence from the corresponding gradient register. Transient local gradient components of environmental feature sequences The weights are calibrated by dynamically updating the adaptive background value of the current step size in real time based on the nonlinear interaction characteristics of the two using a horizontal linear mapping logic. The calculation formula is shown below: Before extracting transient local gradient components, interference needs to be removed from the industrial time-series data sequence. This removal process is based on the principle of linear superposition in discrete-time systems in digital signal processing. By constructing a differential operation loop with frequency selectivity, high-frequency random noise and low-frequency environmental drift are separated from the total measured values. The data standardization preprocessing module inputs the standard time-series data sequence into a sliding window multiplier-adder. The time step length of the time-series sliding window data buffer is set to 8. The sliding window multiplier-adder performs a weighted multiplication-addition operation on the first-order difference quotient of the standard time-series data sequence within the window to obtain the deformation rate gradient. The data arbitration unit calls a preset low-pass digital filter to filter out excess values ​​in the deformation rate gradient. The high-frequency fluctuation component of the 50Hz power frequency interference harmonics is used to obtain the filtered deformation rate index. The data arbitration unit reads the alternating temperature field characteristic time series output by temperature sensors deployed at key load-bearing nodes of the building structure. The temperature sensor measurement accuracy is 0.1℃, and the sampling frequency is 10Hz. Due to the heat conduction delay in the influence of the alternating temperature field on structural deformation, the data arbitration unit introduces a time decay factor based on the first-order delay model of the thermodynamic one-dimensional unsteady heat conduction standard. This factor is used to weight and transform the alternating temperature field characteristic time series, and the temperature change gradient within the fourth cycle preceding the current sampling step is taken as the transient local gradient component of the environmental characteristic sequence at the current step. This eliminates phase asynchrony caused by heat conduction delay, achieving physical and logical decoupling of deformation characteristics from environmental multi-field coupling interference variables. Adjust the weights for the adaptive background value at the current step size. The transient local gradient components of the first-order deformation accumulation sequence at the current step size. This represents the transient local gradient components of the environmental feature sequence at the current step size, with a value of 0.5 serving as the baseline for the weighting coefficient variation; the value of 0.00001 is a pre-set anti-deadlock floating-point constant in the status register, used to establish the division logic protection boundary; subscript Index for the current time step; letter For deformation scalar identification; letters For environmental scalar identification, the mathematical basis for algebraic subtraction of characteristic variables with different physical properties and dimensions in the formula lies in: the transient local gradient components of the first-order deformation accumulation sequence. It is calculated based on the first-order difference quotient of a standard time-series data sequence whose value range has been narrowed to the interval between 0 and 1. Its physical essence has been transformed into a dimensionless scalar of relative deformation rate of change; correspondingly, the transient local gradient components of the environmental feature sequence Before being substituted into the formula, the processor calls the normalization operator to divide the original temperature variation gradient value by the maximum ambient temperature fluctuation range determined in the offline calibration stage, eliminating the Celsius dimension and simultaneously converting it into a dimensionless relative temperature variation scalar in a homogeneous mathematical space, thereby ensuring that the numerator has dimensional equivalence and mathematical consistency in algebraic operations.

[0029] The adaptive background value calibration weight for the current step size is calculated in the adaptive background value calculation module. Subsequently, the data boundary truncation module calibrates the weights based on the adaptive background value. When the value is greater than 0.8, a high-order truncation action is triggered to overwrite the weight value written to the target register with 0.8, and the weight is calibrated with an adaptive background value. When the value is less than 0.2, a low-order truncation action is triggered, overwriting the weight value in the target register with 0.2, and the weight is calibrated by fixing the adaptive background value. The fluctuation range is determined by the adaptive background value calculation module, which calibrates the weights based on the truncated adaptive background value. The weighted values ​​are calculated with the standard time series data sequence, and then reassembled in the background value register to generate a background value sequence adjacent to the mean. The calculation formula is shown below: ,in, For the background value sequence that is close to the mean, Adjust the weights for the adaptive background value after truncation; This is the first-order cumulative deformation sequence for the current step size; For the first-order cumulative deformation sequence of the previous step; subscript Index for the current time step; subscript Index for the previous time step; letter As a deformation scalar identifier, the residual dynamic correction prediction module connects to the adaptive background value calculation module via a channel to extract the nearest mean background value sequence from the background value register. A column vector is constructed and horizontally concatenated with a one-dimensional constant vector of all 1s to build an observation matrix. The transpose product of the observation matrix is ​​calculated to obtain a square matrix. The inverse matrix of this square matrix is ​​then solved. Using this inverse matrix, the transpose of the observation matrix, and the first-order difference vector of the standard time series data sequence, a chain multiplication mapping value is calculated. The development coefficients and driving coefficients of the whitening differential equation are identified online to output a trend prediction sequence. Simultaneously, a residual tracking queue is established in system memory to construct a synchronous difference sequence. The predicted output value and the actual deformed input feature value at the current sampling time are read and compared in real time. The difference between the two is calculated as the predicted residual sequence and stored in the residual tracking queue. The second-order residual correction module extracts the predicted residual from the residual tracking queue. Using the sequence as input parameters, the system fits and predicts the development trend of the residual sequence to output a time-varying deviation compensation value. This value is then superimposed in situ onto the output register to output the deformation trend prediction data. This allows the discrete physical deviation value of the preceding feature gradient to directly constitute the target input for the subsequent background value reconstruction. Simultaneously, the parameter identification result inversely constrains the adaptive boundary of the second-order residual correction module, constructing a strong causal closed loop between the preceding and following computational steps at the memory bus scheduling level. Here, the gradient difference calculated in the preceding step is directly used as the target input for the subsequent background value reconstruction. This means that the gradient difference extracted in the preceding operation is used to constrain and calibrate the adaptive background value in real time by updating the weights in the current time step. This allows for in-situ control and constraint of the current step-size first-order accumulation sequence. Compared with the previous step long first-order cumulative sequence These two physical parameters serve as the target inputs in the arithmetic multiplication and addition weighting proportions of the adjacent mean background value recombination operator. Thus, at the system bus flow level, the preceding gradient feature values, in the form of feedback constraints, indirectly complete the matrix structure reconstruction of the subsequent background value recombination output sequence. The residual dynamic correction prediction module establishes a synchronous tracking difference sequence in parallel in system memory. The second-order residual correction module uses a whitening differential equation system based on a first-order univariate grey residual model to perform a secondary accumulation transformation on the read prediction residual sequence, online fitting the development trend of the prediction residual sequence and outputting the deviation compensation value that evolves with the time step. Specifically, the digital flow path of the parameter identification result inversely constraining the adaptive boundary input is as follows: the microprocessor's arithmetic logic unit will use the online identified development coefficients... The absolute value of the coefficient is used as the input variable of the adaptive control factor to establish an inverse proportional mapping linear operator between the development coefficient and the input cutoff bandwidth; when the development coefficient... When the absolute value of the residual increases and approaches the divergence threshold, the operator synchronously and proportionally narrows the amplitude tolerance level at the input of the second-order residual correction module, dynamically filtering out sudden distortion points caused by transient disturbances in the residual tracking queue. By rigidly shrinking the adaptive boundary of the residual input, it prevents numerical calculation overshoot during subsequent secondary accumulation of the residual correction algorithm. The deformation trend prediction data is loaded as a digital quantity into the output register and directly converted into downlink physical control instructions through the system bus to constrain the mechanical actions of the external entity automated safety protection mechanism. The multi-level safety arbitration output module judges the predicted deformation amount in the deformation trend prediction data through the cascaded numerical comparison circuit inside the microprocessor. When the predicted deformation amount exceeds the fixed orange warning threshold in the system memory for three consecutive control cycles, the multi-level safety arbitration output module sends a second-level deformation over-limit command to the external electrical execution unit through the industrial control network within 14.3ms. The second-level deformation over-limit command is manifested as a pulse width modulation signal containing a specific target voltage amplitude. The electrical execution unit receives the signal and adjusts the connection to the building structure reinforcement lock. The opening of the hydraulic proportional control valve at the fixed mechanism linearly increases the duty cycle of the hydraulic proportional control valve input voltage from the initial 20% to 75%, driving the hydraulic pump to output oil pressure at a rated working pressure of 35MPa to apply reverse mechanical constraint resistance to the key deviation phase state of the building structure, causing the solution matrix to flow downward to the bottom hardware action closed loop. During this process, 14.3ms is the deterministic electronic control cycle delay from the completion of digital over-limit judgment by the safety arbitration output module to the completion of data packet encapsulation and transmission by the control network. In order to overcome the large inertial mechanical execution After the mechanism's fluid dynamics build-up pressure hysteresis, the external hydraulic pump group is in a normal, all-time operating energy storage state during system monitoring and operation. Its pressure output end is connected in parallel with a high-pressure nitrogen accumulator group to pre-establish and maintain a static nominal pressure of 35MPa at the front inlet of the hydraulic proportional control valve. This ensures that when the voltage duty cycle is stepped up to 75%, the high-pressure oil accumulated in the accumulator is released without delay, bypassing the speed and pressure ramp-up phase of the hydraulic pump body. This controls the physical application response of the reverse mechanical resistance within a millisecond-level mechanical window.

[0030] The multi-level safety arbitration output module connects to the residual dynamic correction prediction module via a data bus to receive deformation trend prediction data. It establishes a three-layer cascaded digital over-limit decision logic, comparing the predicted deformation amount in the deformation trend prediction data with fixed yellow warning threshold, orange warning threshold, and red safety limit threshold in the system memory. When the predicted deformation amount exceeds the yellow warning threshold for three consecutive control cycles, it outputs first-level abnormal data; when it exceeds the orange warning threshold, it outputs a second-level deformation over-limit command; and when it exceeds the red safety limit threshold, it outputs a third-level system emergency stop control message. The processed data stream is converted into digital logic control signals that directly regulate the safety status of industrial entities, meeting the requirements of microsecond-level prediction response and deterministic threshold closure arbitration in industrial computing nodes with limited microprocessor computing power and memory bandwidth.

[0031] Example 2: The experimental calculation platform calls upon the offline historical measured deformation time-series dataset collected by a high-precision displacement sensor in the field. The high-precision displacement sensor has a measurement range of 0mm to 50mm, a measurement accuracy of 0.01mm, and a sampling frequency set to 10Hz. A data source with industrial field fluctuation characteristics is established, and the time step length of the time-series sliding window data buffer is [determined]. The value was set to 8 to serve as an optimized working window for balancing computational load and solution accuracy. To simulate a non-ideal industrial environment with multiple electromagnetic interferences and high-frequency micro-vibrations of structures, Gaussian white noise with a signal-to-noise ratio of 25dB and power frequency interference harmonics with a frequency of 50Hz were actively superimposed on the data source to construct an unsteady original input deformation sequence containing strong random fluctuations. During the experiment, multiple test groups were run simultaneously. The test group fully activated the complete algorithm logic consisting of the adaptive background value calculation module, the data boundary truncation module, and the second-order residual correction module, while limiting the adaptive background value calibration weight. The fluctuation range is within a closed interval of 0.2 to 0.8, and the anti-deadlock floating-point constant 0.00001, preset in the state register, is superimposed in the denominator of the division. Control group 1 uses the traditional constant background value operator with a fixed weight coefficient of 0.5, while control group 2 uses adaptive weight adjustment logic without boundary truncation. When the input non-steady original input deformation sequence undergoes a step deformation and the abrupt change amplitude reaches 0.50 mm, the adaptive background value calculation module captures the jump of the external feature gradient in real time, and the adaptive background value calibration weight is dynamically solved in the experimental group. Rapidly shifting towards the boundary, the theoretical weight value is calculated to be 0.89 at the 12th sampling step. The specific calculation formula is shown below: ,in, Adjust the weights for the adaptive background value at the current step size. The transient local gradient components of the first-order deformation accumulation sequence at the current step size. This represents the transient local gradient component of the environmental feature sequence at the current step size.

[0032] The value 0.5 is the baseline for the weighting coefficient variation; the value 0.00001 is a pre-set anti-deadlock floating-point constant in the status register, used to establish the division logic protection boundary; subscript Index for the current time step; letter For deformation scalar identification; letters The environmental scalar is identified; the adaptive background value calibration weight for the current step size is calculated in the adaptive background value calculation module. Subsequently, the data boundary truncation module performs a high-order truncation action, forcibly rewriting the actual output weight to 0.80, and calibrating the weight with a fixed adaptive background value. The fluctuation range is determined by the adaptive background value calculation module, which calibrates the weights based on the truncated adaptive background value. The weighted values ​​are calculated with the standard time series data sequence, and then reassembled in the background value register to generate a background value sequence adjacent to the mean. The specific calculation formula is shown below: ,in, For the background value sequence that is close to the mean, Adjust the weights for the adaptive background value after truncation; This is the first-order cumulative deformation sequence for the current step size; For the first-order cumulative deformation sequence of the previous step; subscript Index for the current time step; subscript Index for the previous time step; letter It is used as a scalar identifier for deformation.

[0033] The high-level control group had its boundary truncation adjusted to 0.95, and the low-level control group had its boundary truncation adjusted to 0.05. Test data showed that when the deformation rate gradient continued to increase and the adaptive weights of the high-level control group entered the range of 0.85 to 0.95, the model's ability to suppress high-frequency power frequency interference harmonics decreased, and the output trend prediction sequence contained digital ripples with an amplitude of 0.041 mm, indicating that the system exhibited noise amplification. This is related to the mathematical mechanism that the high-level weighting coefficient leads to an excessively low weight for the historical smoothing term and the model overfitting random disturbances. As expected, when the deformation signal enters the gradual deformation plateau region and the adaptive weight of the ultra-low limit control group drops to 0.05, a dynamic response lag occurs within the model, increasing the recognition delay for subsequent sudden deformations to 4.2 control cycles. This confirms that the system has entered the response saturation region caused by excessive accumulation of historical information. However, within the optimization working window of 0.2 to 0.8 defined by the experimental group, the aforementioned noise amplification and response lag are controlled within the system's allowable tolerance, maintaining numerical stability throughout the entire process while ensuring nonlinear tracking. Furthermore, a deformation mutation rate gradient system including low, medium, and high intensity levels is introduced to verify the system's quantitative perception capability of industrial disaster risks of different severity levels. Experimental results show that as the deformation mutation rate increases from 0.12 mm / s to 1.54 mm / s, the root mean square error of the prediction residual in the traditional control group increases from 0.021 mm / s to 0.185 mm / s, exhibiting significant overshoot and lag. Meanwhile, the lowest residual sum of squares term in the experimental group remains stable, and its trend prediction sequence consistently maintains high fidelity in following the actual deformation input feature value. The root mean square error of the final prediction residuals under the medium and high intensity gradients stabilized at 0.004 mm, 0.009 mm and 0.015 mm, respectively. The deviation compensation value output by the second-order residual correction module offsets the second-order solution loss caused by time delay. After receiving the deformation trend prediction data, the multi-level safety arbitration output module outputs the corresponding second-level deformation over-limit instruction within 14.3 ms through the three-layer cascaded digital over-limit decision logic, converting the processed data stream into a deterministic digital logic control signal to complete the closed-loop arbitration of the deformation state.

[0034] Example 3: When the system faces the condition of a large bridge structure undergoing continuous and significant nonlinear deviation due to transient impact from high-altitude wind loads, industrial sensors deployed on key load-bearing nodes of the structure acquire industrial time-series data sequences characterizing the deformation state of the building structure with a hardware sampling period of 100ms under memory bus scheduling. The data standardization preprocessing module connects to the time-series data acquisition module through the data bus inside the microprocessor, reads the current industrial time-series data sequence and loads it into the local high-speed data cache, uses a time-domain median filtering algorithm to sort and extract data from a sliding window of 5 consecutive sampling points, removes transient pulse electrical signal anomalies caused by on-site mechanical collisions, and calls the max-min normalization operator to map the filtered time-series values ​​to the standard value range of 0 to 1. The standard time-series data sequence is written in place into the first state register to complete the data cleaning of the input carrier. To quantify the dynamic changes in the nonlinear perturbation behavior of the deformation sequence, the adaptive background value calculation module, drawing on the concept of state shift characteristics in information theory, considers the difference between the standard time-series data sequence of the current sampling step size and its preceding adjacent step size as an impulse response reflecting the degree of abrupt change in the structural deformation rate. The change gradient difference index is defined by calculating the physical cumulative displacement of this impulse response within the time-series sliding window. Within the time-series sliding window data buffer with a time step length limited to 8, the sliding window multiplication and addition value of the deformation rate gradient is calculated. The data arbitration unit extracts the transient local gradient components of the first-order deformation accumulation sequence from the corresponding first state register. Transient local gradient components of environmental feature sequences The adaptive background value calibration weight calculation formula is triggered within the microprocessor's arithmetic logic unit: ,in, Adjust the weights for the adaptive background value at the current step size. The transient local gradient components of the first-order deformation accumulation sequence at the current step size. This represents the transient local gradient components of the environmental feature sequence at the current step size, with a value of 0.5 serving as the baseline for the weighting coefficient variation; the value of 0.00001 is a pre-set anti-deadlock floating-point constant in the status register, used to establish the division logic protection boundary; subscript Index for the current time step; letter For deformation scalar identification; letters For environmental scalar identification.

[0035] When the system reads the transient local gradient components of the first-order deformation accumulation sequence at the current moment from the first state register... The measured value is 0.045, and the transient local gradient components of the environmental feature sequence are read. When the measured value is 0.015, the arithmetic logic unit substitutes the above scalar into the weight calculation formula, calculates the theoretical weight value as 0.50 plus the quotient of 0.030 divided by 0.06001, and writes the weight calculation output value 0.9999 into the target register. Since this weight value exceeds the working window, the data boundary truncation module is triggered and starts the exclusion bottom line control rule. The comparison finds that the value is greater than the preset upper limit threshold of 0.8. Under the hardware shift instruction without relying on external interrupts, a high-bit truncation overwrite action is performed to calibrate the weight of the adaptive background value in the target register. The overlay is rewritten to 0.80. The adaptive background value calculation module calibrates the weights based on the adaptive background value of 0.80 and calculates the weighted value with the standard time series data sequence. The weighted value is then reassembled in the background value register to generate the nearest mean background value sequence of the current time segment. The calculation formula is shown below: ,in, This is a sequence of background values ​​immediately adjacent to the mean, and the unit of this sequence is consistent with that of the first-order cumulative deformation sequence. Adjust the weights for the adaptive background value after truncation; This is the first-order cumulative deformation sequence for the current step size; For the first-order cumulative deformation sequence of the previous step; subscript Index for the current time step; subscript Index for the previous time step; letter The deformation scalar identifier is used; the adaptive boundary mechanism prevents the overshoot of the gray background value solution caused by the weight tending to the extreme value, and the energy dissipation and mass transfer resistance of historical features are offset on the algorithm path.

[0036] The residual dynamic correction prediction module extracts the background value sequence immediately adjacent to the mean value through the internal bus. An 8-dimensional data column vector is constructed. This vector is then horizontally recombined with a one-dimensional constant vector of all ones using a matrix concatenation operator to build the observation matrix. The processor calls a multiplication inverse matrix operation instruction to solve for the square matrix inverse of the product of the observation matrix and its transpose. The resulting inverse matrix, the transpose of the observation matrix, and the first-order difference vector of the standard time-series data sequence in the first state register are multiplied together. The development coefficients and driving coefficients of the whitening differential equation are identified online. Simultaneously with outputting the main fitting trend prediction sequence, the processor allocates a residual difference tracking queue in local static random access memory. It compares the main prediction output value with the actual deformed input feature value at the current sampling time in real time, calculates the physical difference between the two, and stores the predicted residual sequence in the residual difference tracking queue. The second-order residual correction module, as a logical branch of the alternative implementation, reads the predicted residual sequence from the residual difference tracking queue, performs secondary grey modeling on the development trend of the residual sequence, and outputs the deviation compensation value evolving with the time step. This deviation compensation... The value is superimposed in place onto the output register to output deformation trend prediction data, so that the physical residual generated by the preceding module is directly converted into the hard constraint condition of the subsequent calibration loop. The multi-level safety arbitration output module receives the deformation trend prediction data in the output register through the data bus. The cascaded numerical comparison circuit inside the microprocessor performs a three-level cascaded digital over-limit judgment with the predicted deformation amount and the fixed yellow warning threshold of 15.00mm, orange warning threshold of 25.00mm and red safety limit threshold of 35.00mm in the system memory. When the deformation trend prediction data shows that the predicted deformation amount is greater than the orange warning threshold of 25.00mm for three consecutive control cycles, the hardware digital peripheral converts the data stream and sends a digital logic control signal of the second-level deformation over-limit instruction to the external industrial control network within 14.3ms. This drives the on-site automated safety protection mechanism to perform structural reinforcement and locking actions, transforming the logical operation of the original data into a specific engineering on-site decision response. The closed-loop arbitration of the deformation state is thus completed.

[0037] Example 4: When the building structure is in a static condition to be monitored, the system initiates an offline parameter calibration process to establish the operating baseline of the data preprocessing module. The processing unit controls the sensor to continuously sample for 60 seconds without external excitation input, calculates the statistical distribution characteristics of the time series data, and stores them in the fixed memory space. The statistical distribution characteristics are specifically the environmental static noise variance. The above process constructs a noise characteristic model for the current specific deployment environment. The system calculates a reference benchmark for the adaptive adjustment factor based on this model to ensure that the input of the subsequent prediction logic has a definite physical background value attribute.

[0038] When the deformation analysis system enters the online prediction state, the data input port reads the original industrial time-series data sequence. First, calculate the real-time dynamic variance of the sequence within the sliding window. The dynamic variance is logically compared with the noise feature model obtained in the offline calibration stage. If the two satisfy the following relationship: ,in, The real-time dynamic variance within the sliding window. The environmental static noise variance established during the offline calibration phase. This is a preset threshold for determining signal validity. This threshold is a dimensionless coefficient and ranges from 3.0 to 5.0; subscript If the current time step index is used, it indicates that the interference intensity of the current input data exceeds the physical boundary expected by the system design. The processor triggers the data anomaly handling logic, prohibiting the injection of the original data for that time period into the gray model prediction unit and generating an anomaly sampling alarm message. Simultaneously, the system starts the backup data interpolation compensation logic, and performs signal smoothing reconstruction for the anomaly period based on the most recent sample value in the queue that meets the stability criterion. Under the premise of ensuring that the input data quality meets the computational constraints, verifiable deformation trend data is output to the system. This process establishes a deterministic protection closed loop from the physical measurement environment to the algorithm execution logic. Signal validity judgment threshold. The mathematical basis for setting it to 3.0 to 5.0 lies in the variance fluctuation of a stationary random signal in an industrial field within a time series sliding window. The verification principle is as follows: since the system noise under normal conditions has been calibrated, if the real-time dynamic variance within the current sliding window exceeds 3.0 to 5.0 times the variance of the environmental static noise, it means from a mathematical and statistical perspective that the input signal has non-stationary step characteristics where the Gaussian white noise assumption fails. This reflects that there is a random hard jump caused by severe external mechanical impact or sensor disconnection at the physical measurement site. Based on this range, hard truncation gating can accurately block the deadlock caused by abnormal signals in the least squares matrix solution.

[0039] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A grey model prediction system for building structure deformation analysis, characterized in that, include: The time-series data acquisition module is used to acquire industrial time-series data sequences that characterize the deformation state of building structures; The data standardization preprocessing module is connected to the time series data acquisition module through a channel. It is used to filter industrial time series data sequences and standardize and scale them to generate standard time series data sequences and store them in the first state register. The adaptive background value calculation module is connected to the data standardization preprocessing module through a channel. It is used to calculate the difference of the change gradient based on the standard time series data sequence in the first state register to determine the adaptive background value calibration weight. The adaptive background value calibration weight is limited to 0.2 to 0.

8. The weighted value is calculated based on the adaptive background value calibration weight and the standard time series data sequence, and a background value sequence with the nearest mean is generated in the background value register. The residual dynamic correction prediction module is connected to the adaptive background value calculation module through a channel. It is used to extract the nearest mean background value sequence from the background value register to construct the observation matrix. The development coefficient and driving coefficient of the differential equation are obtained by solving the least squares inverse matrix to output the trend prediction sequence. A synchronous difference sequence is established to capture the prediction residual sequence at the current sampling time. The second-order residual correction module calculates the deviation compensation value based on the prediction residual sequence and superimposes the deviation compensation value onto the output register to output the deformation trend prediction data.

2. The grey model prediction system for building structure deformation analysis according to claim 1, characterized in that, In the adaptive background value calculation module, the data boundary truncation module triggers a high-order truncation overwrite to 0.8 when the calculated adaptive background value calibration weight is greater than 0.8, and triggers a low-order truncation overwrite to 0.2 when the adaptive background value calibration weight is less than 0.

2. The module also adds a deadlock prevention floating-point constant of 0.00001 preset in the status register of the adaptive background value calculation module to lock the fluctuation range of the adaptive background value calibration weight.

3. The grey model prediction system for building structure deformation analysis according to claim 1, characterized in that, The data standardization preprocessing module includes the following sub-steps: identifying and removing burst noise terms from the industrial time series data sequence using a time-domain median filtering algorithm; scaling the removed industrial time series data sequence using a max-min normalization algorithm so that the value range of the output standard time series data sequence is within the range of 0 to 1.

4. The grey model prediction system for building structure deformation analysis according to claim 1, characterized in that, The adaptive background value calculation module includes the following sub-steps: calculating the first-order difference quotient of adjacent time step data to obtain the deformation rate gradient; calculating the sliding window multiplication and addition value of the deformation rate gradient to generate the change gradient difference reflecting the degree of temporal fluctuation.

5. A grey model prediction system for building structure deformation analysis according to claim 1, characterized in that, The residual dynamic correction prediction module includes the following sub-steps: establishing a residual tracking queue in system memory, reading the predicted output value and the actual deformation input feature value at the current sampling time in real time; comparing the predicted output value and the actual deformation input feature value, calculating the difference between the two as the predicted residual sequence and storing it in the residual tracking queue.

6. The grey model prediction system for building structure deformation analysis according to claim 1, characterized in that, The residual dynamic correction prediction module includes the following sub-steps: extracting the predicted residual sequence from the residual tracking queue as input parameters and sending it to the second-order residual correction module; fitting the development trend of the predicted residual sequence through the second-order residual correction module and outputting the deviation compensation value that changes over time.

7. A grey model prediction system for building structure deformation analysis according to claim 1, characterized in that, The system also includes a multi-level safety arbitration output module, which is connected to the residual dynamic correction prediction module via a data bus to receive deformation trend prediction data. The multi-level safety arbitration output module is used to establish a three-layer cascaded digital over-limit decision logic, comparing the predicted deformation amount in the deformation trend prediction data with fixed yellow warning threshold, orange warning threshold, and red safety limit threshold in the system memory. When the predicted deformation amount exceeds the yellow warning threshold for three consecutive control cycles, a first-level abnormal data is output; when it exceeds the orange warning threshold, a second-level deformation over-limit command is output; and when it exceeds the red safety limit threshold, a third-level system emergency stop control message is output.

8. A grey model prediction system for building structure deformation analysis according to claim 1, characterized in that, The residual dynamic correction prediction module includes the following sub-steps: extracting the sequence of background values ​​with the nearest mean to form a column vector, and horizontally concatenating it with a one-dimensional all-1 constant vector to construct the observation matrix; The transpose product of the observation matrix is ​​calculated to obtain a square matrix, and the inverse matrix of the square matrix is ​​solved. The multiplication mapping value is calculated using the inverse matrix, the transpose matrix of the observation matrix, and the first-order difference vector of the standard time series data sequence to identify the development coefficients and driving coefficients of the differential equation.

9. A grey model prediction system for building structure deformation analysis according to claim 1, characterized in that, The system constrains the time-series flow through the system memory bus scheduling mechanism. A data bus feedback channel is configured between the deformation trend prediction data output by the residual dynamic correction prediction module and the gradient difference calculated by the adaptive background value calculation module. This allows the gradient difference calculated in the preceding steps to be directly used as the target input for the background value reorganization in the subsequent steps. Furthermore, the parameter identification result inversely constrains the input adaptive boundary of the second-order residual correction module, forming a memory storage architecture for the multi-dimensional feature time-series matrix.

Images