Standard load based force sensor in-situ automatic calibration system

Abstract

The present application relates to the technical field of sensor measurement and automation control, in particular to a force sensor in-situ automatic calibration system based on standard load, comprising: a state sensing step, collecting force sensor and environmental background vibration data, calculating calibration signal-to-noise ratio index and mechanism health index, and combining to generate a system comprehensive state vector; an intelligent game decision step: inputting the vector into a physical information reinforcement learning model, outputting calibration strategy instructions based on the double constraint conditions of sensor drift risk and calibration mechanism life consumption; a closed-loop execution correction step, responding to the instruction to control the mechanism to apply physical standard load, and generating sensor correction parameters based on the loading response data; the present application solves the problem of signal-to-noise ratio inversion under non-stop state, and dynamically balances data accuracy and mechanism maintenance life.

Description

Technical Field

[0001] This invention relates to the field of sensor measurement and automated control technology, specifically to an in-situ automatic calibration system for force sensors based on standard loads. Background Technology

[0002] In the continuous, uninterrupted operation of large industrial equipment, force sensors are subjected to harsh working conditions for extended periods, making them highly susceptible to zero-point drift and resulting in distorted measurement data. To address this issue, existing solutions typically employ offline disassembly calibration or simple online automatic calibration.

[0003] Offline calibration requires forced equipment shutdown, which not only interrupts the production process but also incurs extremely high maintenance costs. Conventional online calibration faces the problem of signal-to-noise ratio inversion under strong background vibration interference, making it difficult to extract effective feature data. More importantly, existing technologies generally lack lifespan management for the calibration actuator itself. Blindly high-frequency calibration can lead to excessive wear or even premature failure of the loading mechanism, while overly conservative low-frequency calibration can lead to the accumulation of sensor drift risk. It is difficult to strike a balance between ensuring data accuracy and extending the lifespan of the mechanism.

[0004] Therefore, how to overcome the downtime limitation in a strong dynamic interference environment and make intelligent decisions to calibrate the machine in order to dynamically solve the dual constraint game between sensor drift correction and calibration mechanism life consumption is an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an in-situ automatic calibration system for force sensors based on standard loads. Specifically, the technical solution of this invention includes:

[0006] Data acquisition unit, status sensing unit, decision control server, and calibration execution unit;

[0007] The data acquisition unit is configured to acquire the force sensor output data and environmental background vibration data of the target device in real time, and send the force sensor output data and environmental background vibration data to the state perception unit.

[0008] The state perception unit is configured to calculate the current calibration signal-to-noise ratio based on the force sensor output data and environmental background vibration data, and to evaluate the health index of the calibration loading mechanism by combining the operating status data fed back by the calibration execution unit. The calibration signal-to-noise ratio index and the health index are then combined to generate a comprehensive system state vector.

[0009] The decision control server is configured to input the system's overall state vector into a preset physical information reinforcement learning model;

[0010] The physical information reinforcement learning model is configured to output calibration strategy instructions based on the dual constraints of sensor drift risk and calibration mechanism life consumption.

[0011] The calibration execution unit is configured to respond to calibration strategy instructions, control the standard load loading mechanism to apply a physical standard load to the force sensor, and collect response data during the loading process to feed back to the state sensing unit.

[0012] The decision control server is also configured to generate sensor correction parameters based on response data and complete in-situ automatic calibration.

[0013] Preferably, the physical information reinforcement learning model is configured with a drift risk loss function and a mechanism lifetime loss function;

[0014] The decision control server is also configured to: calculate the estimated cost of performing a forceful calibration action on the lifespan of the calibration mechanism based on the system's integrated state vector and the mechanism life loss function;

[0015] In response to the detection that the target equipment is in a non-processing and idle state, the sensor drift risk cost caused by maintaining the current state without making corrections is calculated based on the force sensor output data using the drift risk loss function.

[0016] By performing cost minimization optimization calculations using a physical information reinforcement learning model, the optimal decision point is determined between the estimated cost value and the sensor drift risk cost value, in order to generate calibration policy instructions.

[0017] Preferably, the decision control server is also used for:

[0018] The calibrated signal-to-noise ratio is compared with the preset interference threshold.

[0019] In response to the calibration signal-to-noise ratio index being lower than the interference threshold, the current environment is determined to be in a high-risk steady state, and a trend verification strategy instruction is generated as the calibration strategy instruction.

[0020] Alternatively, in response to the calibration signal-to-noise ratio being higher than or equal to the interference threshold, the current environment is determined to be in an ideal steady state, and a full-range accurate calibration command is generated as the calibration strategy command.

[0021] Preferably, the calibration execution unit is further configured to: in response to a trend verification strategy instruction, control the standard load loading mechanism to perform a pulse loading action with an amplitude lower than a preset safety threshold;

[0022] The decision control server is also configured to extract the transient response characteristics of the sensors under pulse loading action;

[0023] The linearity deviation of the force sensor is calculated based on transient response characteristics;

[0024] If the linearity deviation exceeds the preset allowable deviation range, an alarm signal is generated and no full-range parameter correction is performed to reduce the mechanical wear of the calibration loading mechanism.

[0025] Preferably, the state-aware unit is further configured to extract mechanical hysteresis characteristic values ​​and stiffness characteristic values ​​during the loading action execution process from the response data;

[0026] The mechanical hysteresis characteristic value and stiffness characteristic value are compared with the preset set of reference mechanism characteristics;

[0027] If the comparison result exceeds the preset health tolerance range, it is determined that the standard load loading mechanism has a risk of benchmark degradation, and the health index is updated to an abnormal state.

[0028] If the comparison results are within the preset health tolerance range, the standard load loading mechanism benchmark is determined to be reliable, and the health index is updated to normal status.

[0029] Preferably, the decision control server is also used for:

[0030] In response to an abnormal health indicator, the calibration execution unit is locked, new calibration actions are prohibited, and a maintenance request is output.

[0031] If the health indicator is in a normal state, the calibration execution unit is allowed to continue executing subsequent calibration tasks.

[0032] Preferably, the data acquisition unit includes: a high-frequency mechanical sampling module and an environmental spectrum analysis module;

[0033] Among them, the high-frequency mechanical sampling module is used to acquire force sensor output data at a sampling rate higher than the operating frequency of the target device;

[0034] The environmental spectrum analysis module is used to perform modal analysis on environmental background vibration data and identify the resonant mode frequencies of the current operating conditions so that the state sensing unit can eliminate interference in the resonant frequency band.

[0035] Preferably, the decision control server is also used for:

[0036] Calculate the current mechanical measurement error value based on the response data;

[0037] Determine whether the mechanical measurement error value exceeds the preset correction dead zone threshold;

[0038] In response to the mechanical measurement error value exceeding the correction dead zone threshold, the zero-point offset coefficient and sensitivity coefficient of the force sensor are updated using the response data;

[0039] In response to the mechanical measurement error value not exceeding the correction dead zone threshold, the current parameters of the force sensor are kept unchanged.

[0040] Preferably, the decision control server and the state awareness unit are deployed in edge computing nodes or cloud servers;

[0041] The system also includes a remote interactive terminal for visually displaying the system's overall state vector, calibration strategy commands, and sensor correction parameters.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. This invention constructs a signal energy calculation model by introducing an effective lever arm constant and rotating mechanical impedance, achieving closed-loop consistency between torque response and background vibration velocity spectrum in terms of physical dimensions. This effectively solves the calculation distortion problem caused by dimensional differences in traditional signal-to-noise ratio (SNR) evaluation, ensuring the accuracy of identifying the effective calibration window under strong background vibration. By applying a physical information reinforcement learning model, the dual constraints of sensor drift risk and calibration mechanism lifespan consumption are embedded in the decision logic, breaking the limitation that traditional in-situ calibration must be stopped or wait for absolute stillness. It can intelligently capture the mechanical equilibrium window for dynamic calibration while ensuring continuous operation of the main equipment. Through this combination of SNR evaluation based on physical dimension consistency and reinforcement learning decision-making, the system achieves a dynamic balance between data accuracy and mechanism maintenance lifespan, avoiding mechanism damage caused by blind high-frequency calibration or drift and missed detection caused by low-frequency calibration.

[0044] 2. This invention, by configuring a drift risk loss function and a mechanism life loss function, realizes a quantitative game between the physical cost of performing calibration and the risk cost of not performing calibration, constructing an intelligent decision-making mechanism with self-restraint capabilities. By introducing trend extraction and low-pass filtering in drift risk assessment, it effectively prevents high-frequency environmental vibration from being mistakenly integrated as sensor drift, ensuring the physical authenticity of risk cost calculation. Through the joint calculation of the normalized wear coefficient based on fatigue life and the risk weight coefficient based on error impulse, the cost values ​​of different physical dimensions can be compared and optimized under the same benchmark, thereby determining the optimal decision point between the estimated cost and the risk cost. This mechanism effectively avoids equipment self-destruction caused by forced calibration when the calibration mechanism is fragile, and also prevents main equipment accidents caused by excessive conservatism when the drift risk is extremely high, ensuring the optimal maintenance strategy throughout the system's entire life cycle.

[0045] 3. This invention achieves intelligent differentiation between high-risk steady-state and ideal steady-state by comparing the calibration signal-to-noise ratio index with the dynamic interference threshold, and flexibly switches between trend verification strategy and full-range accurate calibration strategy accordingly, maximizing the effective calibration opportunities in complex industrial environments; by using pulse loading action in trend verification and combining it with dynamic gain coefficient based on convolution model to calculate linearity deviation, the sensor status can be quickly detected in a low-damage manner without causing excessive wear to the mechanism; by decoupling and extracting mechanical hysteresis feature values ​​and stiffness feature values ​​from the response data and comparing them with the benchmark, reverse health monitoring of the calibration mechanism itself is realized. Once the benchmark degradation risk is detected, the interlock protection is triggered, effectively preventing erroneous corrections caused by calibration device failure, forming a key line of defense for system reliability;

[0046] 4. This invention utilizes a high-frequency mechanical sampling module in conjunction with an environmental spectrum analysis module. It employs power spectral density estimation and median absolute deviation algorithms to accurately identify the resonant mode frequencies under operating conditions. Furthermore, by constructing a notch filter to eliminate interference in the resonant frequency band within the time domain, it significantly improves the signal-to-noise ratio and purity of the original data. By calculating the mechanical measurement error value and introducing a dead-zone threshold correction mechanism, it avoids high-frequency oscillations of control parameters caused by measurement noise, ensuring the stability of parameter correction. Through linear regression analysis based on the least squares method and inverse function transformation derivation, it achieves accurate joint updates of the sensor's zero-point offset coefficient and sensitivity coefficient, eliminating the residual error impact of gain changes on zero-point propagation. Thus, with the computing power support of edge computing nodes or cloud servers, it achieves high-precision in-situ automatic calibration. Attached Figure Description

[0047] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0048] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0050] Example 1:

[0051] Please see Figure 1 The in-situ automatic calibration system for force sensors based on standard loads includes: a data acquisition unit, a state sensing unit, a decision control server, and a calibration execution unit; wherein, the data acquisition unit is configured to acquire the force sensor output data and environmental background vibration data of the target device in real time, and send the force sensor output data and environmental background vibration data to the state sensing unit;

[0052] The state perception unit is configured to calculate the current calibration signal-to-noise ratio based on the force sensor output data and environmental background vibration data, and to evaluate the health index of the calibration loading mechanism by combining the operating status data fed back by the calibration execution unit. The calibration signal-to-noise ratio index and the health index are then combined to generate a comprehensive system state vector.

[0053] The decision control server is configured to input the system's overall state vector into a preset physical information reinforcement learning model;

[0054] The physical information reinforcement learning model is configured to output calibration strategy instructions based on the dual constraints of sensor drift risk and calibration mechanism life consumption.

[0055] The calibration execution unit is configured to respond to calibration strategy instructions, control the standard load loading mechanism to apply a physical standard load to the force sensor, and collect response data during the loading process to feed back to the state sensing unit; the decision control server is also configured to generate sensor correction parameters based on the response data to complete in-situ automatic calibration.

[0056] This embodiment details the overall architecture and core interaction logic of the system, aiming to solve the problem of zero drift caused by the force sensor due to harsh environment when large industrial equipment is running continuously without stopping, while traditional offline calibration is too costly and online calibration is affected by strong background vibration interference, resulting in an inverted signal-to-noise ratio.

[0057] The system comprises four core interactive units: the data acquisition unit acts as the sensing antenna, responsible for acquiring raw signals with high fidelity and transmitting them via industrial Ethernet or fiber optic bus; the state perception unit acts as the preprocessing center, transforming raw data into state features that can be understood by the AI ​​model, wherein the calibrated signal-to-noise ratio index is derived from the ratio of the spectral energy based on the environmental background vibration data to the estimated response energy of the standard load.

[0058] Specifically, the system collects environmental background vibration data. The velocity power spectral density is obtained by performing a discrete Fourier transform. For processing velocity power spectral density A rectangular window function is introduced. The mathematical definition of a rectangular window function is: it takes a value of 1 when the absolute value of the independent variable is less than or equal to 0.5, and 0 otherwise. Its calculation formula is:

[0059]

[0060] in, This is a rectangular window function used to define and remove the frequency band window containing the calibration signal in the subsequent integral calculation of ambient background noise energy. The system uses a preset calibration frequency, which is set by the user based on the inherent frequency of the equipment spindle. To prevent the calibration signal itself from being misinterpreted as noise, the system is set to use the calibration frequency. The center bandwidth The frequency is typically taken as 0.5-2.0Hz. The background noise energy is calculated using an integral formula, which is as follows:

[0061]

[0062] in, This represents the upper limit of the maximum analysis frequency for spectral analysis of environmental background vibration data. For frequency integral variables; regarding the calculation of signal energy, to eliminate inconsistencies and dimensional differences with the physical model benchmark in drift risk assessment, the system introduces a preset effective lever arm constant. This constant is obtained from the calibration of the mechanism's geometry; the signal energy is calculated by mapping the preset amplitude of the standard load to the current dynamic boundary conditions of the physical model, specifically using the following formula:

[0063]

[0064] in, To calibrate the angular frequency, satisfy the following conditions: ; This section introduces The physical basis is as follows: to ensure the closed-loop consistency of physical dimensions and to correct the applicability error of the linear impedance model, this embodiment... Clearly defined as the resistance of rotating machinery, its physical dimension is the ratio of torque to angular velocity, that is... Therefore, the numerator of the formula needs to be constructed as torque. With lever arm The product of, i.e. ,dimension This makes the molecular weight dimensions With the denominator impedance dimension After division, the calculation Strictly reduced to the dimension of linear velocity Thus, the theoretical tangential response velocity at the end of the lever arm can be accurately characterized.

[0065] in, Input amplitude for standard load. For calibrating frequency This refers to the mechanical impedance modulus of the equipment at a specific circular frequency under standard load, which is determined by analyzing the real-time acquired spindle torque signal. With speed signal Perform synchronous sliding window FFT transform to extract the calibration frequency. The amplitude components at the point are calculated and the ratio is calculated. This allows us to obtain the square of the theoretical response velocity amplitude, which includes the real-time stiffness characteristics of the equipment, as the signal energy. This ensures the consistency of the signal-to-noise ratio assessment across physical dimensions, i.e., the velocity square ratio. The logarithm of this ratio is then taken. As a calibrated signal-to-noise ratio indicator;

[0066] The health index is evaluated by combining the operating status data fed back by the calibrated execution unit, such as motor current and oil pressure response time. Through a weighted normalization method, the system quantifies the health level according to the following formula. To maintain consistency of notation throughout the text, this indicator will be uniformly denoted as [insert notation here] in subsequent formulas involving time variations. To ensure the traceability of variables, the calculation formula is as follows:

[0067]

[0068] in, This is the measured peak current when the calibration mechanism performs its operation. To calibrate the rated current when the loading mechanism performs its operation, The measured hydraulic pressure settling time is defined as the time span required from the moment the calibration command is issued until the hydraulic cylinder pressure reaches 90% of the preset target value. As the baseline response time, The preset weighting coefficients are typically determined experimentally based on the actuator's sensitivity to current overload and response hysteresis, and recommended values ​​are selected. And satisfy ;

[0069] The formula directly outputs the normalized dimensionless value. However, considering that in the event of a serious equipment malfunction, such as stalled rotor or oil circuit blockage, the deviation term may be greater than 1, leading to... The calculation result is negative, which triggers a sign reversal of the lifecycle penalty term, meaning the more severe the fault, the higher the reward. This embodiment explicitly specifies that a non-negative truncation operator is applied to the output of this formula, i.e., the final health value is [value missing]. To ensure that it is strictly implemented Within the interval; the two are concatenated to generate the system's comprehensive state vector; specifically, the system constructs a feature vector; here, given the signal-to-noise ratio... Typically, it's a larger value between 10dB and 60dB; just input it directly. The function causes the output to constantly approach 1, making the neural network unable to perceive changes in the environment. Therefore, the system pre-programs... The formula for amplitude scaling is as follows:

[0070]

[0071] in The preset normalization constant is set to the maximum ideal signal-to-noise ratio designed for the system. This value is determined by the upper limit of the sensor's dynamic range, for example, 100dB, and is used to map the input value to... The linear sensitivity region of a function; for The normalization function is expressed as follows: The aim is to map physical features of different dimensions to the [0,1] interval to ensure the numerical stability of the neural network input; the decision control server, as the brain of the system, has a built-in physical information reinforcement learning model, which is a deep neural network that embeds physical constraints into the reward function. Specifically, it adopts a proximal policy optimization or deep deterministic policy gradient algorithm architecture. Its input layer dimension is consistent with the dimension of the system's comprehensive state vector. The hidden layer consists of multiple fully connected layers and ReLU activation functions. The output layer is mapped to a discrete or continuous calibration action space.

[0072] The model receives a state vector as input and engages in a game under the dual constraints of sensor drift risk and the lifespan consumption of the standard load loading mechanism (also referred to as the calibration loading mechanism or calibration mechanism throughout the text). Specifically, to address the problem of converting physical constraints into a computable reward function, enabling those skilled in the art to reproduce the core decision-making logic, this embodiment explicitly defines an immediate reward function based on a physical penalty term. Its mathematical expression is:

[0073]

[0074] in, This is the system's overall state vector at the current moment. The calibration action command output for the decision at the current moment. For the current action The preset mechanical loss equivalent, Where 1 represents the full-range calibration action, corresponding to quantity 1.0, and 0 represents the standby action, corresponding to quantity 0; This is an indicator of the organization's health at the current moment. To prevent local minima where the denominator is zero, for example... , The preset penalty weighting coefficient is used to balance the proportion of institutional losses and drift risk in the total reward. For example, take... , This is the timestamp of the last successful full-range calibration and parameter update; this item utilizes the inverse proportional function property to ensure that when the mechanism's health... During descent, the penalty value for performing the calibration action increases sharply in a hyperbolic manner, thus forming a physical soft limit with fault protection at the algorithm level;

[0075] The second item is the sensor drift risk item, in which Defined as the absolute value of the deviation between the measured value of the sensor after low-pass filtering and the theoretical reference value, i.e. This quantifies the time. The degree of instantaneous drift;

[0076] The model performs policy optimization based on this formula, specifically through an Actor-Critic network architecture, where the Critic network utilizes an immediate reward function. The value function or advantage function of the current state is calculated. Based on this value feedback, the Actor network updates the network weight parameters using gradient descent to adjust the action output probability in order to maximize the long-term cumulative reward. This establishes a mapping relationship from the input state vector to the action instruction, and finally outputs the optimal calibration policy instruction.

[0077] The calibration execution unit acts as the executor, responding to commands to control the standard load loading mechanism to apply a physical standard load to the force sensor, and simultaneously collects response data feedback to generate correction parameters;

[0078] This embodiment introduces a physical information reinforcement learning model, breaking the limitation that traditional in-situ calibration must be stopped or wait for absolute stillness. In a strong dynamic interference environment, the system can intelligently identify fleeting mechanical balance windows for calibration. While ensuring the continuous operation of the main equipment, it dynamically balances the accuracy of sensor data with the maintenance life of the standard load loading mechanism itself, avoiding damage to the mechanism caused by blind high-frequency calibration or drift and missed detection caused by low-frequency calibration.

[0079] Example 2:

[0080] The physical information reinforcement learning model is configured with a drift risk loss function and a mechanism life loss function. The decision control server is also configured to: calculate the estimated cost of performing a force calibration action on the calibration mechanism life consumption based on the system's comprehensive state vector and the mechanism life loss function; in response to detecting that the target equipment is in a non-processing idle state, calculate the sensor drift risk cost caused by maintaining the current state without correction based on the force sensor output data and the drift risk loss function; and perform cost minimization optimization calculations through the physical information reinforcement learning model to determine the optimal decision point between the estimated cost value and the sensor drift risk cost value in order to generate calibration strategy instructions.

[0081] This embodiment elaborates on the decision-making logic of the physical information reinforcement learning model in the decision control server, particularly the calculation process for solving the bi-objective trade-off calibration game problem. Based on the health index in the system's comprehensive state vector, the estimated cost value is calculated using the mechanism life loss function, which reflects the marginal cost of mechanism wear under standard load. As a preferred implementation, this embodiment uses an exponential function instead of the inverse proportional function in Embodiment 1 to obtain smoother gradient descent characteristics. The calculation formula is as follows:

[0082]

[0083] in, The wear coefficient is derived from the basic wear coefficient of the hardware material of the standard load-bearing mechanism. To eliminate the arbitrariness of parameter selection and ensure physical consistency, this coefficient is based on the rated fatigue life in the mechanism's manufacturer's manual. The unit is times. The calculation formula is as follows:

[0084]

[0085] Where, constant This refers to the benchmark of one million cycles defined in the ISO 281 standard for rated life of rolling bearings. Lifetime, used as a normalization denominator, is used to uniformly map mechanisms with different lifetime orders to a standard reference frame. This is a normalized reference constant, derived from statistical analysis of historical accelerated life test data of the same type of loading mechanism, and typically ranges from [value range missing]. The specific value is determined by the shape parameter of the Weibull distribution of historical lifetime data. The aim is to map the physical loss of a single action to a dimensionless value of the same order of magnitude as the risk cost. The physical meaning is the relative weight of the basic loss of a single action in the cost function, thereby ensuring the calculated... Dimensionless numerical values ​​to match Dimensions;

[0086] For hydraulic servo loading mechanisms, if ,but ; Derived from a preset constant, its physical meaning is a lifespan decline sensitivity factor, used to adjust the penalty for declining health, and its value usually ranges from 2.0 to 5.0; The health index originates from the state sensing unit, and its physical meaning is the current physical integrity of the standard load loading mechanism.

[0087] Assuming the target equipment is in a non-processing, unloaded state, based on the current force sensor output data, especially the long-term trend term, the drift risk loss function is used to calculate the sensor drift risk cost, quantifying the potential consequences of error accumulation due to lack of correction. To avoid misintegrating high-frequency environmental vibrations as sensor drift, a trend extraction step is introduced before the integration calculation. To ensure the consistency and comparability of the physical dimensions of the cost function, the calculation formula is as follows:

[0088]

[0089] in, Cutoff frequency The low-pass filter operator, in this embodiment, is a second-order Butterworth filter with a cutoff frequency of [missing information]. The frequency is set to 0.1Hz to filter out high-frequency vibrations from the environment, thus retaining only the slowly changing DC bias drift term; The risk weighting coefficient is derived from the safety level setting of the main equipment, and its specific value is determined by the maximum cumulative drift error impulse allowed by the system. That is, the upper limit of the integral of force error with respect to time, which is determined by working backwards, and its calculation formula is:

[0090]

[0091] Here, the constant 100 is a dimensionless percentage mapping factor, designed to account for the allowable error impulse. The corresponding risk value is normalized to 100, i.e., 100% risk, making its numerical magnitude similar to... To prevent gradient vanishing or dominance bias during neural network training due to excessively large differences in the values ​​of the two cost terms, it is important to keep them on the same order of magnitude, i.e., in the range of 0-100. The specific value is determined based on the safety specifications in the target equipment's machining process manual. It is generally set as the maximum cumulative error integral limit allowed by the equipment's Safety Integrity Level (SIL). For example, for precision grinding, based on the sensitivity analysis of workpiece surface quality to cutting force fluctuations, the value is set, and its calculation formula is:

[0092]

[0093] In the formula, and These represent units of physical quantity, specifically Newton and second;

[0094] It is configured to have The physical dimensions are intended to cancel out the effects of the integral term. The dimension, in physical terms, represents the penalty multiplier for drift risk. By mapping the dimension of this coefficient, the physical impulse error is transformed into a dimensionless risk cost value, thus... Capable of relating to the dimensionless processing of mechanism life loss Direct algebraically weighted comparisons are performed under the same benchmark;

[0095] The physical basis for integrating the absolute value of the error over time here is that in continuous mechanical monitoring systems, the risk of drift depends not only on the magnitude of the instantaneous error, but also on the cumulative effect of the error over time, i.e., the error impulse; for example, in precision machining scenarios, The total amount of errors is directly proportional to the cumulative deviation of workpiece dimensions or the amount of material removed; this integral form quantifies the potential risk of process defects into a physical cumulative amount, thus representing the risk cost. It provides a clear physical leverage point, enabling it to engage in an equivalent game with the physical wear and tear costs of the mechanism;

[0096] Derived from the current measured values ​​of the sensor, the physical meaning is mechanical data including potential drift;

[0097] The theoretical force values ​​are derived from the predictions of the physical model, and their physical meaning is the mechanical reference under ideal conditions. This physical model is pre-constructed based on the dynamic equations of the mechanical transmission chain of the target device. To avoid black-box descriptions caused by the term "based on," in the scenario where the target device is driven by a motor and the force sensor is located at the end of the transmission chain, the following dynamic observer equations are constructed for real-time solution. The calculation formula is as follows:

[0098]

[0099] in, This is the real-time input current of the main drive motor. This refers to the motor speed. The torque constant of the motor. The equivalent rotational inertia of the transmission system. The system's viscous damping coefficient is... The length of the lever arm; for the noise-sensitive angular acceleration term in the formula. To prevent direct differential from introducing high-frequency quantization noise, Distortion is addressed by a specially configured second-order low-pass differential filter with a transfer function designed as follows:

[0100]

[0101] in, Set to 5-8 times the motor control bandwidth. A value of 0.707 is used to obtain a smooth acceleration estimate with controllable phase delay; the above parameters All of these are known physical constants pre-stored in the system, and are configured to use the current, speed and control commands of the main drive motor of the target device as input variables to calculate the theoretical load reference value in real time without relying on the force sensor.

[0102] It should be noted that the formula uses the fixed lever arm of the calibration mechanism. The torque-force conversion theoretical model is only valid when the main machine is in a non-machining, no-load state, i.e., when the spindle load rate is below a preset cutting threshold, such as 5%. In this case, the external cutting torque is theoretically zero; any non-zero torque will be invalid. The calculated values ​​all characterize the nonlinear drift or disturbance within the system; the decision control server is configured to activate the drift risk penalty only when this idle state is detected. Cumulative calculations are used to avoid discrepancies between the actual cutting force arm during machining and... Inconsistency leads to model misjudgment.

[0103] It originates from the time recorded by the system when the last full-range accurate calibration was performed and the parameter correction was confirmed to be complete, which is the starting point of the drift error accumulation period;

[0104] The physical information reinforcement learning model performs cost minimization optimization computation, specifically by minimizing the estimated cost value. Compared with the cost of sensor drift risk The weighted sum and negatives are used as immediate rewards. ,in For dimensionless normalized weighting factors, satisfying This is used to adjust the relative importance of the decision model to the protection of the mechanism's lifespan and the maintenance of measurement accuracy; the network parameters are updated using the policy gradient algorithm to maximize the cumulative expected return, thereby determining the optimal decision point between the estimated cost value and the sensor drift risk cost value, so as to generate calibration policy instructions; when the drift risk is extremely high and the standard load loading mechanism is healthy, the model tends to generate strong calibration instructions, otherwise it chooses to postpone calibration.

[0105] This embodiment achieves a self-restraining intelligent decision-making mechanism by quantifying the physical cost of performing calibration and the risk cost of not performing calibration. This mechanism effectively avoids equipment self-destruction caused by forced calibration when the standard load loading mechanism is vulnerable, and also prevents main equipment accidents caused by excessive conservatism when the drift risk is extremely high, ensuring the optimal maintenance strategy throughout the system's entire life cycle.

[0106] Example 3:

[0107] In this embodiment, the decision control server is further configured to: compare the calibrated signal-to-noise ratio (SNR) index with a preset interference threshold; in response to the SNR index being lower than the interference threshold, determine that the current environment is in a high-risk steady state and generate a trend verification strategy instruction as a calibration strategy instruction; or, in response to the SNR index being higher than or equal to the interference threshold, determine that the current environment is in an ideal steady state and generate a full-range accurate calibration instruction as a calibration strategy instruction.

[0108] This embodiment specifically illustrates the logic of the decision control server dynamically switching calibration modes based on the level of environmental interference; the system obtains the calibration signal-to-noise ratio index calculated by the state perception unit and compares it with a preset interference threshold. It is not a fixed constant, but is dynamically generated by the environmental noise baseline statistically analyzed by the system during equipment idle periods. Its calculation formula is as follows:

[0109]

[0110] in, and These represent the mean and standard deviation of the background noise within the previous quiet time window, respectively. The confidence coefficient is recommended to be 3.0 to ensure that the threshold can adaptively fluctuate with the operating condition reference. This threshold is the lowest signal-to-noise ratio boundary to ensure the reliability of the full-range calibration data. In response to the calibration signal-to-noise ratio index being lower than the interference threshold, the system determines that the current environment is in a high-risk steady state, i.e., a high-risk interference state, to avoid confusion with the physical steady state. At this time, the environmental noise is too large, and precise calibration may introduce noise errors. Therefore, a trend verification strategy instruction is generated to perform a low-precision probe action to confirm the sensor status.

[0111] Conversely, in response to the calibration signal-to-noise ratio being higher than or equal to the interference threshold, the system determines that the current environment is in an ideal steady state, i.e., an ideal low-interference state, which is suitable for high-precision parameter correction, generates full-range accurate calibration instructions, and instructs the execution unit to perform multi-level load loading to obtain a complete mechanical characteristic curve.

[0112] This embodiment addresses the drawback of blindly pursuing high precision under non-stationary operating conditions. By distinguishing between high-risk steady state and ideal steady state, the system proactively degrades to a survival mode that only verifies trends when the environment is harsh, and switches to a correction mode for precise calibration when the environment is favorable. This dynamic switching mechanism maximizes the use of the effective calibration window and significantly improves the system's adaptability and data reliability in complex industrial environments.

[0113] Example 4:

[0114] The calibration execution unit is also configured to: respond to the trend verification strategy command and control the standard load loading mechanism to perform a pulse loading action with an amplitude lower than a preset safety threshold; the decision control server is also configured to: extract the transient response characteristics of the sensor under the pulse loading action; calculate the linearity deviation value of the force sensor based on the transient response characteristics; if the linearity deviation value exceeds the preset allowable deviation range, generate an alarm signal and do not perform full-range parameter correction to reduce the mechanical wear of the calibration loading mechanism.

[0115] This embodiment is a detailed description of the trend verification strategy command, aiming to illustrate how to detect the sensor status without causing excessive wear on the mechanism. When responding to the trend verification strategy command, the calibration execution unit controls the standard load loading mechanism to perform a pulse loading action, i.e., applying a mechanical pulse with an extremely short duration, such as 10ms-50ms, and an amplitude below a preset safety threshold, such as 20% of full scale. The decision control server extracts the transient response characteristics under this pulse loading action from the data fed back by the sensor, and calculates the linearity deviation value of the force sensor accordingly. The calculation formula is as follows:

[0116]

[0117] in, The measured peak response is derived from the sensor output. Given that this step is performed under high background noise conditions, i.e., the high-risk steady state in Example 3, directly reading the peak value is highly susceptible to interference. Therefore, the system employs matched filtering technology to extract this value: a pre-stored standard unit impulse response sequence is used. As a template, a cross-correlation operation is performed with the real-time acquired sensor data stream, and the global maximum value of the cross-correlation result is taken as the template. The signal-to-noise ratio is significantly improved by utilizing the lack of correlation between random noise and deterministic impulse response;

[0118] The standard pulse load amplitude is derived from the control command setting.

[0119] This represents the theoretical dynamic gain coefficient under the current pulse width; to ensure dimensional consistency, this embodiment explicitly specifies the control command sequence. The theoretical waveform sequence of the target standard load, whose amplitude is in force and numerically equal to... The coefficient is obtained by using the system's pre-stored standard unit impulse response sequence of the sensor. This sequence was obtained through prior high-frequency system identification experiments, such as using pseudo-random binary sequence excitation or high-frequency step response analysis, and physically corresponds to the Dirac sequence. The system response of the function, combined with the currently issued pulse control command sequence. Performing discrete convolution operation is as follows:

[0120]

[0121] in, For the pre-stored standard unit impulse response sequence The theoretical response sequence is obtained from the data length, and its calculation formula is as follows:

[0122]

[0123] Among them, due to and All are measured in units of force, and thus the following is obtained. Let be a dimensionless dynamic transfer efficiency coefficient, such that... The calculation results keep the force dimension unchanged, giving the deviation formula clear physical executability, thereby eliminating the amplitude attenuation error caused by the pulse duration being shorter than the sensor settling time;

[0124] Response to linearity deviation value Exceeding the preset allowable deviation range, for example The system determines that the sensor may have been physically damaged or severely drifted, generates an alarm signal, and forces the system not to perform full-range parameter correction.

[0125] This embodiment introduces a dynamic gain coefficient based on a convolution model. This enables precise quantification of short-time pulse responses, avoids misjudgments caused by dynamic attenuation, and constructs a low-damage, high-confidence sensor health check mechanism.

[0126] Example 5:

[0127] The state perception unit is also configured to: extract mechanical hysteresis characteristic values ​​and stiffness characteristic values ​​during the loading action execution process from the response data; compare the mechanical hysteresis characteristic values ​​and stiffness characteristic values ​​with a preset set of reference mechanism characteristics; if the comparison result exceeds the preset health tolerance range, it is determined that the standard load loading mechanism has a reference degradation risk, and the health index is updated to an abnormal state; if the comparison result is within the preset health tolerance range, it is determined that the standard load loading mechanism reference is reliable, and the health index is updated to a normal state.

[0128] This embodiment specifically illustrates the reverse calibration logic of the state perception unit performing benchmark health self-sensing on the standard load loading mechanism itself;

[0129] The system achieves feature decoupling from the response data of the calibration process, extracting mechanical hysteresis and stiffness feature values ​​during the loading action. The response data includes a time-synchronized sequence of force sensor output force values. Actuation displacement sequence of standard load loading mechanism Stiffness eigenvalues The least squares method was used to fit and calculate the data of the loaded linear segment:

[0130]

[0131] in, This represents the total number of data points collected in the linear loading segment. The first Displacement and force values ​​at each sampling point; These are the arithmetic mean of the displacement and force values ​​in this data segment; and the mechanical hysteresis characteristic value. The physical area enclosed by the force-displacement closed loop is calculated using the trapezoidal numerical integration formula as follows:

[0132]

[0133] in, The total number of sampling points for a complete loading-unloading cycle is used to accurately quantify the clearance wear and elastic decay of the standard load loading mechanism. The former reflects the non-overlap of force paths during loading and unloading, while the latter reflects the standard load loading mechanism's ability to resist deformation. The measured characteristic values ​​are compared with a preset set of characteristic features of the benchmark mechanism to calculate the relative deviation index. The calculation formula is as follows:

[0134]

[0135]

[0136] in, It is an index representing the relative deviation of stiffness characteristics. It is a relative deviation index for mechanical hysteresis characteristics;

[0137] This set is derived from ideal measurement values ​​after the institution leaves the factory or undergoes major repairs; if the comparison results are... Exceeding the preset health tolerance range, for example If the standard load loading mechanism is deemed to have a risk of reference degradation, then the health index will be determined. Force update to an abnormal state value, such as setting The zeroing operation here is the same as in Example 2. The continuous definition is not contradictory, but rather regarded as a domain of health. The lower boundary state in the context; physically, this means that when a mechanism experiences structural degradation at the baseline feature level, its usability instantly collapses to zero;

[0138] This unified definition method ensures that in the cost function of Example 2, It can trigger an infinite penalty value through the denominator, thus perfectly matching the decision-making mechanism of continuous degradation in mathematical logic. This state change will be directly identified by the decision control server and trigger the lockout protection logic of Example 6. If the comparison result is within the preset health tolerance range, the standard load loading mechanism benchmark is determined to be reliable, and the health index is updated to the normal state.

[0139] This embodiment solves the benchmark trust crisis problem existing in the in-situ calibration system; by using sensor data to monitor the physical characteristics of the standard load loading mechanism in reverse, the system can detect the degradation of the calibration device itself in a timely manner, ensuring the rigor of the calibration process, preventing erroneous corrections caused by calibrator failure, and forming a second line of defense for system reliability.

[0140] Example 6:

[0141] In this embodiment, the decision control server is also used to: lock the calibration execution unit, prohibit the execution of new calibration actions, and output an organization maintenance request in response to an abnormal health indicator; and allow the calibration execution unit to continue executing subsequent calibration tasks in response to a normal health indicator.

[0142] This embodiment further illustrates the interlocking mechanism based on health indicators, serving as a barrier for system security. The decision control server continuously monitors the status of the health indicators. In response to an indicator being determined to be in an abnormal state, the system immediately triggers the abnormal locking logic, physically cutting off the control signal of the calibration execution unit, forcibly prohibiting the execution of new calibration actions, and outputting a maintenance request to the operation and maintenance terminal. In response to an indicator being in a normal state, the system executes the normal release logic, allowing the calibration execution unit to continue responding to subsequent calibration strategy instructions.

[0143] This embodiment constructs a key interlocking mechanism to prevent system self-destruction; when the calibration mechanism shows signs of wear or failure, continuing to perform calibration will not only fail to obtain correct results, but will also accelerate the damage to the mechanism; through forced locking, this mechanism achieves active protection of the calibration hardware and isolates potential hazards outside the control loop, avoiding the spread of faults.

[0144] Example 7:

[0145] The data acquisition unit includes a high-frequency mechanical sampling module and an environmental spectrum analysis module. The high-frequency mechanical sampling module is used to acquire force sensor output data at a sampling rate higher than the operating frequency of the target device. The environmental spectrum analysis module is used to perform modal analysis on the environmental background vibration data and identify the resonant mode frequencies of the current working condition so that the state sensing unit can eliminate interference in the resonant frequency band.

[0146] This embodiment details the hardware configuration and signal processing logic of the data acquisition unit, aiming to improve the purity of the raw data. The unit includes a high-frequency mechanical sampling module configured to acquire force sensor output data at a sampling rate higher than the target device's operating frequency, for example, more than 10 times the base frequency of the main device. It also includes an environmental spectrum analysis module for performing the following specific modal analysis and interference removal algorithms:

[0147] PSD estimation: based on ambient background vibration data Calculation of power spectral density using the Welch method , use length Hamming window and 50% overlap;

[0148] Resonance Mode Recognition: Traversal Identify frequency points that meet the criteria for prominent peak values. The judgment condition is:

[0149]

[0150] in, and These are the mean and standard deviation of the current spectral baseline within the effective frequency band after removing the DC component and low-frequency drift noise in the 0-5Hz range. The specific calculation uses the median absolute deviation estimation method. This is to prevent high-energy resonance peaks from shifting the statistical benchmark and causing missed detections, thereby dynamically locking the resonance interference frequency point under the current operating conditions.

[0151] Notch filter construction: targeting the identified resonant frequency Construct a second-order IIR notch filter with the following transfer function: Defined as:

[0152]

[0153] in, For normalized angular frequency, This refers to the sampling frequency used by the system when performing high-frequency sampling of environmental background vibration data. For preset bandwidth The determined pole radius, The filter is applied to the raw mechanical data stream to accurately remove interference energy in the resonant frequency band in the time domain, ensuring the signal-to-noise ratio of the calibration data.

[0154] Example 8:

[0155] The decision control server is also used to: calculate the current mechanical measurement error value based on the response data; and determine whether the mechanical measurement error value exceeds the preset correction dead zone threshold.

[0156] When the mechanical measurement error exceeds the correction dead zone threshold, the zero-point offset coefficient and sensitivity coefficient of the force sensor are updated using the response data; when the mechanical measurement error does not exceed the correction dead zone threshold, the current parameters of the force sensor are kept unchanged.

[0157] This embodiment illustrates the specific calculation logic of the decision control server during the parameter correction phase; the system is based on the response data of the calibration stable section and the measured values ​​of the sensors. Compared with standard load reference value The current mechanical measurement error value is calculated, using the root mean square error as the quantification index. The calculation formula is as follows:

[0158]

[0159] in, This represents the total number of sampling data points contained within the stable calibration segment;

[0160] judge Does it exceed the preset dead zone threshold? ,For example ,in The constant for the correction dead zone determination threshold set for the system;

[0161] In response to The system executes the parameter update algorithm as follows: 1. Regression analysis: Constructing a linear model The measured gain of the current physical system is calculated using the least squares method. and measured bias The calculation formula is as follows:

[0162]

[0163] Coefficient correction: based on measured parameters Inverse correction of the sensor's internal sensitivity coefficient and zero-point offset coefficient The updated formula is:

[0164]

[0165] This calculation formula is strictly based on The inverse function transform derivation corrects the neglect of gain variation. The residual error that may result from the influence of zero-point offset propagation is addressed through this update, resulting in a corrected sensor output. Able to approximate the actual load again ;

[0166] In response to The system determines that the current error is within the allowable measurement noise range and keeps the current parameters unchanged to prevent the control parameters from oscillating at high frequencies near the dead zone.

[0167] Example 9:

[0168] In this embodiment, the decision control server and the state perception unit are deployed in edge computing nodes or cloud servers; the system also includes a remote interactive terminal for visually displaying the system's comprehensive state vector, calibration strategy instructions, and sensor correction parameters.

[0169] This embodiment illustrates the system's deployment architecture and human-computer interaction method. The decision control server and state perception unit are deployed in the form of software containers on edge computing nodes or cloud servers in the industrial field. This architecture provides powerful computing support for running complex physical information reinforcement learning models. In addition, the system also includes remote interactive terminals, such as the host computer screen in the central control room, which are used to visualize the system's comprehensive state vector, including the signal-to-noise ratio heatmap and the trend of the mechanism's health, while also displaying the historical records and current values ​​of calibration strategy commands and sensor correction parameters.

[0170] This embodiment solves the problem of complex game models' dependence on computing resources by deploying at the edge or in the cloud; at the same time, the remote interactive terminal provides a transparent monitoring window, enabling operation and maintenance personnel to grasp the system's decision-making process and the organization's physical health status in real time, realizing the interpretability and visualization of black-box algorithms, and enhancing the trust in human-machine collaboration.

[0171] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An in-situ automatic calibration system for force sensors based on standard loads, characterized in that, include: Data acquisition unit, status sensing unit, decision control server, and calibration execution unit; The data acquisition unit is configured to acquire the force sensor output data and environmental background vibration data of the target device in real time, and send the force sensor output data and the environmental background vibration data to the state sensing unit. The state perception unit is configured to calculate the current calibration signal-to-noise ratio index based on the force sensor output data and the environmental background vibration data, and to evaluate the health index of the calibration loading mechanism by combining the operating status data fed back by the calibration execution unit, and to combine the calibration signal-to-noise ratio index and the health index to generate a comprehensive system state vector. The decision control server is configured to input the system's comprehensive state vector into a preset physical information reinforcement learning model; The physical information reinforcement learning model is configured to output calibration strategy instructions based on the dual constraints of sensor drift risk and calibration mechanism life consumption. The calibration execution unit is configured to respond to the calibration strategy command by controlling the standard load loading mechanism to apply a physical standard load to the force sensor and collecting response data during the loading process to feed back to the state sensing unit. The decision control server is also configured to generate sensor correction parameters based on the response data and complete in-situ automatic calibration.

2. The in-situ automatic calibration system for force sensors based on standard loads according to claim 1, characterized in that, The physical information reinforcement learning model is configured with a drift risk loss function and an mechanism lifetime loss function; The decision control server is also configured to: calculate the estimated cost of performing a forceful calibration action on the lifespan of the calibration mechanism based on the system's comprehensive state vector and using the mechanism lifespan loss function; In response to the detection that the target equipment is in a non-processing and idle state, the sensor drift risk cost caused by maintaining the current state without correction is calculated based on the force sensor output data and using the drift risk loss function. The physical information reinforcement learning model performs cost minimization optimization calculations to determine the optimal decision point between the estimated cost value and the sensor drift risk cost value, thereby generating the calibration strategy instruction.

3. The in-situ automatic calibration system for force sensors based on standard loads according to claim 1, characterized in that, The decision control server is also used for: The calibrated signal-to-noise ratio index is compared with a preset interference threshold; In response to the calibration signal-to-noise ratio index being lower than the interference threshold, the current environment is determined to be in a high-risk steady state, and a trend verification strategy instruction is generated as the calibration strategy instruction; Alternatively, in response to the calibration signal-to-noise ratio being higher than or equal to the interference threshold, it is determined that the current environment is in an ideal steady state, and a full-range accurate calibration instruction is generated as the calibration strategy instruction.

4. The in-situ automatic calibration system for force sensors based on standard loads according to claim 3, characterized in that, The calibration execution unit is further configured to: in response to the trend verification strategy instruction, control the standard load loading mechanism to perform a pulse loading action with an amplitude lower than a preset safety threshold; The decision control server is also configured to: extract the transient response features of the sensor under the pulse loading action; The linearity deviation value of the force sensor is calculated based on the transient response characteristics; If the linearity deviation value exceeds the preset allowable deviation range, an alarm signal is generated and no full-range parameter correction is performed to reduce the mechanical wear of the calibration loading mechanism.

5. The in-situ automatic calibration system for force sensors based on standard loads according to claim 1, characterized in that, The state sensing unit is also configured to extract mechanical hysteresis characteristic values ​​and stiffness characteristic values ​​from the response data during the loading action execution process. The mechanical hysteresis characteristic value and the stiffness characteristic value are compared with a preset set of reference mechanism characteristics; If the comparison result exceeds the preset health tolerance range, it is determined that the standard load loading mechanism has a risk of benchmark degradation, and the health index is updated to an abnormal state. If the comparison result is within the preset health tolerance range, the standard load loading mechanism benchmark is determined to be reliable, and the health index is updated to a normal state.

6. The in-situ automatic calibration system for force sensors based on standard loads according to claim 5, characterized in that, The decision control server is also used for: In response to the health indicator being in an abnormal state, the calibration execution unit is locked, new calibration actions are prohibited, and a maintenance request is output. In response to the health indicator being in a normal state, the calibration execution unit is allowed to continue executing subsequent calibration tasks.

7. The in-situ automatic calibration system for force sensors based on standard loads according to claim 1, characterized in that, The data acquisition unit includes: a high-frequency mechanical sampling module and an environmental spectrum analysis module; The high-frequency mechanical sampling module is used to acquire the force sensor output data at a sampling rate higher than the operating frequency of the target device. The environmental spectrum analysis module is used to perform modal analysis on the environmental background vibration data and identify the resonant mode frequencies of the current working condition so that the state sensing unit can eliminate interference in the resonant frequency band.

8. The in-situ automatic calibration system for force sensors based on standard loads according to claim 1, characterized in that, The decision control server is also used for: Calculate the current mechanical measurement error value based on the response data; Determine whether the mechanical measurement error value exceeds a preset correction dead zone threshold; In response to the mechanical measurement error value exceeding the correction dead zone threshold, the zero-point bias coefficient and sensitivity coefficient of the force sensor are updated using the response data; In response to the mechanical measurement error value not exceeding the correction dead zone threshold, the current parameters of the force sensor are kept unchanged.

9. The in-situ automatic calibration system for force sensors based on standard loads according to claim 1, characterized in that, The decision control server and the state awareness unit are deployed on edge computing nodes or cloud servers; The system also includes a remote interactive terminal for visually displaying the system's overall state vector, the calibration strategy instructions, and the sensor correction parameters.

Images