Sensor production data recognition method, device and medium applying artificial intelligence

Abstract

The application provides a sensor production data recognition method and device applying artificial intelligence, equipment and medium, a single-dimensional attenuation test sequence in sensor testing is acquired, a deep neural network recognition model embedding a physical degradation mechanism is constructed, the model contains a physical constraint cycle layer, a forgetting gate bias of the physical constraint cycle layer is initialized according to a sensitive material relaxation time value, a hidden state cycle connection weight is constructed as a toeplitz arrangement structure with a sub-diagonal element monotonically decaying with a transmission time step length, a physical consistent hidden state sequence is generated by inputting a sequence into the layer, a decay tendency descriptor is generated by analyzing a stability decay trend, and a long-term stability recognition result is obtained by mapping according to a preset rule. The application realizes automation and high-precision discrimination of sensor long-term stability evaluation.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method, device and medium for identifying sensor production data using artificial intelligence. Background Technology

[0002] Sensor production data identification is a key technology for analyzing response signals acquired during sensor production testing to assess their long-term stability. Existing techniques typically involve directly acquiring the decay response sequence of the sensor's sensitive material and judging batch stability by extracting simple statistical features such as peak values ​​and decay times, or by using empirically set decay rate thresholds. However, because the degradation process of sensor sensitive materials is influenced by complex physicochemical mechanisms, the actual response signal often exhibits a nonlinear and time-varying decay pattern. Existing methods struggle to accurately capture the inherent degradation evolution trend of the material from limited-duration test data, leading to significant discrepancies between long-term stability predictions and actual conditions, thus affecting the reliability of sensor batch quality screening. Summary of the Invention

[0003] This invention provides a method, device, and medium for identifying sensor production data using artificial intelligence.

[0004] In a first aspect, embodiments of the present invention provide a sensor production data identification method using artificial intelligence, comprising: A one-dimensional attenuation test sequence is formed by arranging the response signals obtained through the data acquisition system during the sensor production and testing process in chronological order. The one-dimensional attenuation test sequence consists of multiple response values ​​arranged in chronological order according to the acquisition timestamps. Each response value corresponds to the state of the sensor's sensitive material at a specific time point. A deep neural network recognition model embedding the physical degradation mechanism is constructed. The deep neural network recognition model includes a physical constraint recurrent layer. The forget gate bias term of the physical constraint recurrent layer is initialized based on the relaxation time value of the sensor's sensitive material. The weight structure used to connect the hidden states in the physical constraint recurrent layer is constructed as a Toplitz arrangement structure, and the values ​​of each diagonal element in the Toplitz arrangement structure are set to decrease monotonically with the increase of the propagation time step. The single-dimensional decay test sequence is input into the physical constraint recurrent layer step by step. The physical constraint recurrent layer performs physical mechanism-constrained recurrent processing on the input of each time step by initializing the forget gate bias term and connecting the weights of the Toplitz permutation structure, generating the hidden state vector of each time step, and combining them in time sequence to obtain the physically consistent hidden state sequence. Stability decay trend analysis is performed on physically consistent hidden state sequences. The decay path of the hidden state vector in the physically consistent hidden state sequence changes with time, and a decay tendency descriptor is generated. Using the degradation tendency descriptor as input, the degradation tendency descriptor is subjected to category mapping processing according to the preset long-term stability judgment rule to obtain the long-term stability identification result corresponding to the sensor production batch.

[0005] Secondly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the above-described method.

[0006] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the methods described above.

[0007] This invention constructs a deep neural network model embedded with physical degradation mechanisms. Utilizing a forget gate bias term initialized based on the relaxation time of the sensor's sensitive material and a monotonically decaying weight structure arranged in a Toplitz pattern, the cyclic processing strictly adheres to the physical laws of material stress relaxation, generating a physically consistent hidden state sequence. This effectively suppresses non-physical oscillations or drifts that may occur in purely data-driven models, significantly improving the fidelity of the hidden state sequence in representing the real degradation process. Based on this, by analyzing the stability degradation trend of the physically consistent hidden state sequence and extracting degradation tendency descriptors, combined with pre-defined long-term stability judgment rules for category mapping, the long-term stability category of sensor production batches can be automatically and accurately identified. This avoids the subjective biases and limitations of traditional methods that rely on human experience or simple threshold judgments. This method organically integrates physical mechanisms with deep learning, achieving efficient prediction of sensor long-term stability without destructive testing. It provides reliable technical support for batch quality assessment in sensor production, helping to reduce testing costs and improve production screening efficiency. Attached Figure Description

[0008] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the specification, serve to explain the technical solutions of the present invention.

[0009] Figure 1 This is a schematic diagram illustrating the implementation process of a sensor production data recognition method using artificial intelligence, provided in an embodiment of the present invention.

[0010] Figure 2 and Figure 3 This is a logical schematic diagram of a sensor production data identification method using artificial intelligence, provided in an embodiment of the present invention.

[0011] Figure 4This is a schematic diagram of the hardware entity of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0012] This invention provides a method for identifying sensor production data using artificial intelligence, which can be executed by a processor of a computer device. The computer device can refer to a server, laptop, tablet, desktop computer, smart TV, mobile device (e.g., mobile phone, portable video player, personal digital assistant, dedicated messaging device, portable gaming device), or any other device with data processing capabilities.

[0013] Please combine Figure 1 and Figure 2 Referring to the present invention, an embodiment of a sensor production data identification method using artificial intelligence includes the following steps S100~S500: Step S100: Obtain a one-dimensional attenuation test sequence formed by arranging the response signals obtained through the data acquisition system during the sensor production and testing process in chronological order. The one-dimensional attenuation test sequence consists of multiple response values ​​arranged in chronological order according to the acquisition timestamps. Each response value corresponds to the state of the sensor's sensitive material at a specific time point.

[0014] The sensor production testing process refers to a series of standardized tests performed under controlled environmental conditions before a sensor leaves the factory, after the core sensing element has been prepared, electrode wire bonding, and encapsulation and curing processes have been completed. These tests aim to assess the long-term drift characteristics and stability of the sensor's output signal. Controlled environmental conditions include, for example, a temperature environment provided by a precision constant temperature chamber or a temperature environment that varies according to a preset program, a temperature and humidity combined environment provided by a constant temperature and humidity chamber, and a clean electromagnetic environment free from external electromagnetic interference provided by an electromagnetic shielding box. The data acquisition system is a complete set of signal acquisition, conversion, time stamping, and storage devices deployed at each production and testing station. For example, it can consist of a sensor interface adapter unit, a signal conditioning circuit, an analog-to-digital converter, a clock synchronization module, and a data aggregation and storage computer.

[0015] The sensor interface adapter unit provides a unified flexible contact or pluggable electrical connection for sensors with different package types and pin outputs, leading out the weak analog electrical signals generated by the sensor's sensitive material under the test environment in differential form to the signal conditioning circuit. The signal conditioning circuit consists of a three-stage cascaded structure of a pre-amplifier, a programmable gain amplifier, and an anti-aliasing low-pass filter. The pre-amplifier provides a high common-mode rejection ratio to convert the differential signal into a single-ended signal and completes the first-stage fixed-gain amplification. The programmable gain amplifier dynamically adjusts the second-stage gain factor according to the input range requirements of the analog-to-digital converter through a digital control interface to ensure that the signal amplitude fully utilizes the resolution of the analog-to-digital converter. The anti-aliasing low-pass filter adopts a Bessel or Butterworth filter topology and sets the cutoff frequency to less than half of the sampling frequency to completely filter out high-frequency noise and interference that may cause spectral aliasing.

[0016] Triggered by the sampling clock pulse provided by the clock synchronization module, the analog-to-digital converter (ADC) performs sample-and-hold and quantization encoding on the anti-aliasing filtered analog signal at a preset and fixed sampling frequency, converting the continuous analog signal waveform into discrete digital response values. The clock synchronization module incorporates a high-stability temperature-controlled crystal oscillator and a binary counter. The reference frequency generated by the temperature-controlled crystal oscillator is divided to generate the sampling clock pulse, which simultaneously drives the counter to accumulate. At the beginning of each ADC cycle, the instantaneous count value of the counter is latched and used as the acquisition timestamp for that sampling. The acquisition timestamp accurately records the absolute or relative moment when the response value is digitized.

[0017] The data aggregation and storage computer interacts with the analog-to-digital converter and clock synchronization module through a high-speed industrial bus. It reads the response value generated by each sampling and its corresponding acquisition timestamp into a memory ring buffer in a direct memory access manner, and then writes them in batches into a time-series database file in solid-state non-volatile memory in ascending order of acquisition timestamps. This forms a single-dimensional attenuation test sequence that is strictly arranged in the order of acquisition timestamps.

[0018] The response signal refers to the change in electrical output of a sensor's sensitive material under test conditions, caused by the spontaneous evolution of its internal physical state over time. This evolution includes mechanisms such as the gradual release of residual stress, slow decay of polarization intensity, redistribution of defect concentration, and filling and emptying of carrier traps. These mechanisms all lead to a unidirectional drift in the macroscopic transduction characteristics of the sensitive material. The response value is the digitally encoded value output by the analog-to-digital converter after sampling and quantizing the response signal. It has a definite mapping relationship with the physical state of the sensor's sensitive material at a specific time point, which can be described by the sensor's static calibration curve.

[0019] For piezoresistive sensors, the response value is approximately linearly related to the residual stress state of the sensitive film; for piezoelectric sensors, the response value is approximately proportional to the residual polarization intensity of the sensitive ceramic; for capacitive sensors, the response value corresponds to the capacitance value determined by the dielectric constant and geometry of the sensitive dielectric layer. The one-dimensional attenuation test sequence is a one-dimensional ordered set of values ​​formed by sequentially connecting the response values ​​obtained throughout all the acquisition cycles in ascending order of acquisition timestamps. Each element in the sequence uniquely corresponds to an acquisition time step, and the element value is the response value at that time step. The sequence as a whole exhibits an attenuation pattern where the initial response value gradually decreases over time and eventually approaches asymptotic stability. This attenuation pattern directly reflects the physical degradation process of the sensor's sensitive material at the electrical signal output level.

[0020] Step S200: Construct a deep neural network recognition model embedded with a physical degradation mechanism. The deep neural network recognition model includes a physical constraint recurrent layer. The forget gate bias term of the physical constraint recurrent layer is initialized based on the relaxation time value of the sensor's sensitive material. The weight structure used to connect the hidden states in the physical constraint recurrent layer is constructed as a Toplitz arrangement structure. The values ​​of each diagonal element in the Toplitz arrangement structure are set to decrease monotonically with the increase of the propagation time step.

[0021] Please refer to Figure 3 This is an implementation logic for step S200 in an embodiment of the present invention. Specifically, it can be summarized as the following steps S210 to S260: Step S210: Obtain stress relaxation test records of the sensor-sensitive material collected by constant strain test. The stress relaxation test records contain a stress value sequence arranged in logarithmic time axis order. Each stress value in the stress value sequence corresponds to a sampling time point. The starting stress value of the stress value sequence is the initial stress of the material at the moment when constant strain is applied, and the ending stress value is the residual stress after long-term relaxation.

[0022] The constant strain test is a mechanical testing method specifically designed to study the stress relaxation behavior of materials, performed on standard geometric specimens of sensor-sensitive materials using a computer-controlled precision universal testing machine. The testing machine is equipped with a high-rigidity loading frame, a servo motor-driven crossbeam, a high-precision displacement grating ruler, high-resolution force sensors mounted coaxially in series, and a closed-loop servo control system. At the start of the test, the machine's crossbeam applies compressive or tensile displacement to the specimen at a constant rate. The displacement grating ruler provides real-time feedback on the specimen's current deformation. When the specimen's strain reaches the preset target constant strain, the servo control system immediately switches from displacement rate control mode to strain hold control mode. In strain hold control mode, the servo system uses the feedback signal from the displacement grating ruler as the closed-loop control input and adjusts the servo motor torque output in real-time using a proportional-integral-derivative (PID) control algorithm to counteract the elastic recovery trend of the specimen caused by stress relaxation, ensuring that the macroscopic strain of the specimen remains within a very small tolerance range near the target value throughout the entire test period. The force sensor is fixed in the loading chain and connected in series with the specimen on the same axis. It directly senses the axial load generated by elastic strain inside the specimen and converts it into an analog voltage signal proportional to the load. This voltage signal is then amplified and filtered by the dynamic strain gauge and digitally acquired by the analog-to-digital converter.

[0023] The stress relaxation test record is a complete test data archive automatically generated by the testing machine control and data acquisition software. This archive contains two parts: test metadata and a stress value sequence. The metadata records test conditions such as the target constant strain, test temperature, and specimen dimensions. The stress value sequence is a series of stress measurements collected using a logarithmic time axis sampling strategy and arranged in ascending order on the logarithmic time coordinate. The logarithmic time axis is a non-linear time scale construction method. It uses the start of the test as the zero point and the logarithm of actual physical time as the axis scale, selecting sampling time points at equal intervals on the logarithmic time coordinate. The advantage of using a logarithmic time axis is that it can densely cover the early stage of rapid stress relaxation with a limited number of sampling points across a total test duration spanning several orders of magnitude, while sparsely covering the extremely slow stress change at the end of the test, thus comprehensively and efficiently capturing the dynamic characteristics of the entire stress relaxation process. Each stress value in the stress value sequence corresponds to a unique logarithmic time coordinate sampling time point, which records both the logarithmic time coordinate value and an actual physical time value. The initial stress value of the stress value sequence is defined as the stress value recorded at the first effective sampling time point after the testing machine servo system switches from loading mode to strain holding mode. This value reflects the combined stress level of the elastic and viscoelastic mechanical responses generated by the material when it is forced to withstand a constant strain. The final stress value of the stress value sequence is defined as the stress value recorded at the preset termination logarithmic time coordinate when the test continues. At this moment, the stress relaxation rate of the material has decreased to a minimum level, and the further change of stress over time is negligible. This stress value can be approximated as the quasi-equilibrium residual internal stress level of the material under this constant strain.

[0024] Step S220: Perform relaxation inflection point location processing on the stress value sequence. Search for the change range of stress decrease rate in reverse from the tail to the head of the stress value sequence. When the ratio of stress decrease rate in adjacent logarithmic time intervals exceeds the preset rate change limit, mark the boundary point of the time interval as the inflection point and extract the logarithmic time coordinates corresponding to the inflection point as the inflection time value.

[0025] The relaxation inflection point localization process is an automatic identification workflow for inflection points in material relaxation dynamics based on a logarithmic time domain stress decay rate abrupt change detection algorithm. The stress decay rate is defined as the quotient obtained by dividing the difference in stress values ​​at the two ends of an interval formed by two adjacent logarithmic time sampling points in a stress value sequence by the difference in the logarithmic time coordinates at the two ends of that interval. This quotient represents the absolute decrease in stress value per unit logarithmic time interval, and its dimension is stress units. The magnitude of the stress decay rate change refers to the relative difference between the stress decay rate of the latter interval and the stress decay rate of the former interval in two consecutive adjacent logarithmic time intervals. In this step, the ratio of the stress decay rate of the latter interval to the stress decay rate of the former interval is used as the specific measure of the magnitude of change. A ratio greater than 1 indicates that stress decay is accelerating, and a ratio less than 1 indicates that stress decay is decelerating.

[0026] The reverse search is a traversal strategy for locating relaxation inflection points. Its execution logic starts from the logarithmic time interval at the end of the stress value sequence, treating this interval as the next interval under investigation, and the immediately preceding interval as the previous interval. The ratio of the stress decrease rate between the two intervals is calculated, and it is determined whether it exceeds a preset rate change limit. If it does not exceed the limit, the investigation window is moved forward one interval (i.e., one interval step towards the beginning of the sequence), and the above calculation and judgment are repeated. If, during a certain step, the ratio of the stress decrease rate between the two intervals exceeds the preset rate change limit for the first time, the search is immediately stopped, and the shared intermediate boundary sampling point between the two intervals is determined as the relaxation inflection point. The preset rate change limit is a critical rate ratio threshold determined in advance through statistical analysis of a large amount of stress relaxation test data of similar materials. The principle for setting this threshold is to reliably distinguish the transition region between the rapid and slow decay kinetic stages of material stress relaxation in a statistical sense. When the ratio of the stress decrease rate between adjacent intervals exceeds this threshold, it is considered that the physical mechanism of stress decay has undergone a structural transformation from fast to slow. The inflection point is a specific sampling time point in the stress value sequence corresponding to the relaxation transition. This point belongs to both the preceding and following logarithmic time intervals and is a shared point on their time boundaries. The logarithmic time coordinate value corresponding to this point is the inflection time value.

[0027] Step S230: Determine the logarithmic time coordinate value corresponding to the inflection point as the relaxation time value of the sensor-sensitive material. The relaxation time value represents the position of the characteristic inflection point of the material stress relaxation process on the logarithmic time axis. It is taken as a dimensionless logarithmic coordinate value as the characteristic time scale of the physical degradation process of the sensor-sensitive material.

[0028] The relaxation time value of the sensor's sensitive material is numerically equivalent to the transition time value extracted in step S220 through relaxation transition point location processing; both are dimensionless logarithmic time coordinate values. The relaxation time value, in its physical essence, characterizes the critical transition moment on the logarithmic time axis at which the internal stress relaxation dynamics of this batch of sensor sensitive materials transitions from an early stage of rapid stress release to a late stage of extremely slow stress release after being subjected to constant strain excitation. Since the logarithmic time coordinate itself is a dimensionless value obtained by taking the logarithm of actual physical time, the relaxation time value naturally possesses dimensionless characteristics, and its magnitude is not affected by the time unit system used. A larger relaxation time value means that the material maintains a higher stress relaxation rate over a longer time span, its rapid relaxation phase lasts longer, and the non-equilibrium relaxation mechanism of the material's microstructure can continue to function over a longer time scale; a smaller relaxation time value means that the material ends rapid relaxation earlier and enters the slow relaxation tail stage earlier, its internal stress completes most of the release in a short time, and the remaining residual stress will slowly decay over a long period. The relaxation time value serves as a characteristic timescale of the physical degradation process of the sensor's sensitive material, providing deterministic physical constraint information from the material itself for the parameter configuration of the subsequent physical constraint loop layer. This enables different parameter initialization states of the network model to adaptively match the inherent relaxation dynamics differences of different batches of sensitive materials.

[0029] Step S240: Assign a value to the forget gate bias term based on the relaxation time value. According to the preset bias determination relationship, map the reciprocal of the relaxation time value to the initial bias assignment value. The bias determination relationship makes the initial bias assignment value and the relaxation time value change in opposite directions. The larger the relaxation time value, the smaller the initial bias assignment value, so as to give the forget gate an activation tendency that is adapted to the relaxation speed of the material at the initial time step.

[0030] In one implementation, step S240 may specifically include the following steps S241 to S246: Step S241: Extract the initial stress value and the inflection point stress value from the stress relaxation test record, calculate the difference between the initial stress value and the inflection point stress value, divide the difference by the initial stress value to obtain the stress relaxation amplitude ratio, which characterizes the degree of stress attenuation of the material when it reaches the inflection point.

[0031] The initial stress value is the stress value recorded at the first sampling point in the stress value sequence. This value can be directly read from the stress relaxation test record data structure obtained in step S210 by index or key value. The inflection point stress value refers to the stress value corresponding to the sampling time point marked as the inflection point in step S220. This value can be obtained by accurately searching or linearly interpolating the logarithmic time coordinate value corresponding to the inflection point in the stress value sequence. The difference between the initial stress value and the inflection point stress value reflects the absolute decrease in the total stress amplitude released from the sensor-sensitive material during the period from the start of the test to the inflection point of the relaxation dynamics. The stress relaxation amplitude ratio is defined as the ratio of this difference to the initial stress value. The specific calculation process is to subtract the inflection point stress value from the initial stress value to obtain the stress drop, and then divide the stress drop by the initial stress value. The result is a dimensionless real number between zero and 1. This real number quantitatively expresses the proportion of the initial stress in the material that has decayed and disappeared at the moment the relaxation inflection point arrives. The closer the ratio is to 1, the more the initial stress has been relaxed and released before the inflection point, and the more thorough the stress attenuation is during the rapid relaxation phase of the material. The closer the ratio is to zero, the more the initial stress has been released when the inflection point is reached, and a large amount of stress will gradually attenuate during the slow relaxation phase after the inflection point.

[0032] Step S242: Using the logarithmic time coordinate value corresponding to the inflection point as the logarithmic time span, calculate the ratio of the stress relaxation amplitude to the logarithmic time span to obtain the average relaxation rate descriptor. The average relaxation rate descriptor characterizes the degree of stress decay per unit logarithmic time on the logarithmic time scale.

[0033] The logarithmic time span is the logarithmic time coordinate value corresponding to the inflection point. Since the logarithmic time coordinate at the start of the experiment is usually set to zero or a minimum value close to zero, the logarithmic time coordinate value at the inflection point is numerically equal to the total logarithmic time length spanned from the start of the experiment to the relaxation inflection point. The calculation process of the average relaxation rate descriptor is to use the stress relaxation amplitude ratio obtained in step S241 as the dividend and the logarithmic time span as the divisor, perform a division operation, and obtain the quotient. The physical meaning of this quotient is the proportion of stress attenuation averaged over each unit logarithmic time interval within the complete logarithmic time interval from the start to the inflection point. Its value comprehensively reflects the speed of the macroscopic average relaxation rate of the material during the rapid relaxation phase. A larger average relaxation rate descriptor value indicates that the material has completed a larger share of stress release within a unit logarithmic time interval, and the rapid relaxation phase is more intense and efficient; a smaller average relaxation rate descriptor value indicates that the material has only completed a smaller share of stress release within a unit logarithmic time interval, and the rapid relaxation phase is relatively mild and lasts longer.

[0034] Step S243: Normalize the average relaxation rate descriptor. Using the maximum and minimum observed average relaxation rates among similar materials of the sensor's sensitive material as reference limits, linearly compress the average relaxation rate descriptor to the normalized interval to generate a normalized relaxation rate descriptor.

[0035] The term "similar materials" in sensor terminology refers specifically to a collection of materials that share the same main crystalline phase chemical composition, crystal system, doping formulation, and similar grain size distribution and density level as the sensing materials used in this batch of sensors. The maximum and minimum observed average relaxation rates are pre-determined reference values ​​for this similar material system. These values ​​are obtained by collecting stress relaxation test data from a large number of independent batches, strictly repeating steps S241 to S242 for each batch, and then taking the maximum and minimum values ​​from the average relaxation rate descriptions of all batches. These two values ​​together define the empirical distribution range of the average relaxation rate for this similar material system. Normalization employs a linear minimum-maximum normalization algorithm. The calculation process involves first subtracting the minimum observed average relaxation rate from the current batch's average relaxation rate description to obtain the difference, then subtracting the minimum observed average relaxation rate from the maximum observed average relaxation rate to obtain the range. Dividing the difference by the range yields the linearly compressed normalized relaxation rate descriptor. The normalized relaxation rate descriptor falls within a preset standard normalization interval. The lower bound of this interval is zero, which corresponds to the slowest average relaxation rate among similar materials, and the upper bound is 1, which corresponds to the fastest average relaxation rate among similar materials. The specific value of the descriptor indicates the relative position of the average relaxation rate of the current batch of sensitive materials in the similar material system.

[0036] In this embodiment, the maximum and minimum observed values ​​of the average relaxation rate for the same material system, barium titanate-based piezoresistive ceramic, have been statistically determined and stored in the material parameter database. The maximum and minimum observed values ​​of the current batch average relaxation rate descriptor obtained in step S242 are read from the database, and linear normalization calculation is performed. The generated normalized relaxation rate descriptor is a dimensionless value in the interval between zero and 1.

[0037] Step S244: Input the normalized relaxation rate descriptor into the preset bias mapping relationship. The bias mapping relationship is constructed such that the larger the normalized relaxation rate descriptor is, the smaller the output bias amount is. When the normalized relaxation rate descriptor is close to the upper bound of the normalization interval, the output bias amount approaches the preset lower bound of the bias. When it is close to the lower bound, the output bias amount approaches the preset upper bound of the bias. Thus, the initial bias amount without limiting is obtained.

[0038] The preset bias mapping is a monotonically decreasing continuous function with the normalized relaxation rate descriptor as the independent variable and the unlimited initial bias as the dependent variable. This function uniquely maps each possible value of the normalized relaxation rate descriptor in the interval from 0 to 1 to a corresponding value within the bias range. The function can be constructed using a linear mapping: the unlimited initial bias equals the preset upper bound minus the product of the absolute values ​​of the differences between the normalized relaxation rate descriptor and the preset upper and lower bounds. The preset upper and lower bounds are two numerical limits artificially set based on the reasonable initialization range of the forget gate bias term in the physical constraint loop layer. The preset upper bound allows for a larger value to be assigned to the bias term, and the preset lower bound allows for a smaller value. When the normalized relaxation rate descriptor is close to zero, meaning the average relaxation rate is relatively slow among similar materials, the bias output of the bias mapping relationship approaches the preset upper bound of the bias; when the normalized relaxation rate descriptor is close to 1, meaning the average relaxation rate is relatively fast among similar materials, the bias output of the bias mapping relationship approaches the preset lower bound of the bias. Through this mapping, materials with slower average relaxation rates and milder physical degradation processes will receive larger forget gate biases, while materials with faster average relaxation rates and more severe physical degradation processes will receive smaller forget gate biases. Thus, differential adaptation to the degradation kinetics of different materials is achieved at the initial assignment level of the bias parameters. The output value of the mapping function is the unlimited initial bias value, which has not yet undergone truncation processing for range limiting protection.

[0039] In this embodiment, the normalized relaxation rate descriptor generated in step S243 is substituted into the preset bias mapping relationship. The closer the independent variable is to the upper bound, the closer the output is to the preset lower bound. The closer the independent variable is to the lower bound, the closer the output is to the preset upper bound. The value of the mapped output is the initial value of the unlimited bias.

[0040] Step S245: Perform range limiting processing on the unlimited initial bias amount. If the unlimited initial bias amount exceeds the preset upper limit of the bias tolerance, it is truncated to the upper limit value. If it is lower than the preset lower limit of the bias tolerance, it is truncated to the lower limit value, thus obtaining the limited bias amount.

[0041] The preset bias upper bound and preset bias lower bound are separate sets of parameter boundaries from the preset bias upper bound and preset bias lower bound in step S244, specifically used to safely limit the bias amount output by the mapping function. The upper bound is set as the maximum safe bias value that the forget gate bias term can accept without causing the S-shaped growth curve activation function output to remain completely saturated. The lower bound is set as the minimum safe bias value that will not cause the activation function output to remain completely cut off. The range limiting process first compares the unlimited initial bias amount with the preset bias upper bound. If the unlimited initial bias amount is greater than the upper bound, the value is directly replaced with the upper bound value. Then, the comparison result is compared with the preset bias lower bound. If the unlimited initial bias amount is less than the lower bound, the value is replaced with the lower bound value. If the unlimited initial bias amount falls between the upper and lower bounds, the original value remains unchanged. The limiting bias obtained through this process ensures that the initial value of the forget gate bias will not exceed the safe range under any material and test conditions, thus ensuring the basic stability of the forget gate gating behavior when the physical constraint loop layer is running on different production batches of sensors.

[0042] Step S246: The limiting bias amount is used as the initial bias assignment amount of the forget gate bias term and stored in the bias initialization interface of the physical constraint loop layer so as to complete the physical constraint setting of the forget gate bias before the loop processing starts.

[0043] The bias initialization interface of the physical constraint recurrent layer is the parameter configuration entry point exposed by the deep neural network recognition model during the parameter initialization phase. This interface accepts a scalar value and writes it into the memory address or tensor storage area of ​​the corresponding forget gate bias term in the computational graph structure inside the physical constraint recurrent layer. The write operation is completed before the model performs any forward recurrent propagation calculations, and the written value is the limiting bias amount generated in step S245. After the write is completed, the forget gate bias term is assigned a specific initial value in the network computational graph that carries the relaxation time characteristics of the sensitive material of this batch of sensors. During subsequent model training, the forget gate bias term can be updated according to the backpropagation of the loss function gradient, but the initial value already contains strong constraints of physical laws, making the parameter search starting point located in a reasonable region that matches the material properties. Alternatively, the forget gate bias term can be frozen to a fixed value and not participate in training updates. In this case, the physical constraints maintain a completely deterministic effect throughout the entire model lifecycle.

[0044] Step S250: Based on the relaxation time value and the stress decay law characterized by the stress relaxation test record over time, determine the decay multiplier sequence of each diagonal element value in the Toplitz arrangement structure as the transmission time step increases. Multiply each decay multiplier with the main diagonal element value to obtain a unified value for all elements on the corresponding diagonal, so that the values ​​of adjacent diagonal elements show a monotonically decreasing trend as the offset increases.

[0045] In one implementation, step S250 may specifically include the following steps S251 to S256: Step S251: Extract the stress value segment from the turning point time value to the end of the sequence from the stress relaxation test record as the relaxation late decay sequence. The relaxation late decay sequence reflects the stress decay process after the material enters the slow relaxation stage.

[0046] The truncation operation uses the inflection time value determined in step S220 as the dividing line. From all sampling points included in the stress relaxation test records, all sampling points whose logarithmic time coordinates are greater than or equal to the inflection time value and less than or equal to the logarithmic time coordinate at the end of the sequence are selected. These sampling points are then recombined in ascending order of their original logarithmic time coordinates to form a late-stage relaxation decay sequence. This sequence only includes stress evolution data after the material stress relaxation process has passed the inflection point and entered the slow relaxation tail phase, excluding data from the rapid relaxation stage before the inflection point. Since the late-stage relaxation decay sequence describes the long-term stress decay behavior of the material after the relaxation kinetic transition is complete, its stress change pattern is more singular and stable, making it suitable as the data basis for extracting long-term power-law decay characteristic parameters of the material.

[0047] Step S252: Perform logarithmic domain transformation on the relaxation late decay sequence. First, divide each stress value in the sequence by the preset reference stress value to obtain the stress ratio. Then, take the logarithm of the stress ratio to generate dimensionless logarithmic stress values. The time coordinates of the relaxation late decay sequence have been given in the stress relaxation test record according to the logarithmic time axis. Combine the dimensionless logarithmic stress values ​​with the logarithmic time coordinates to form a logarithmic stress-logarithmic time point set.

[0048] The preset reference stress value can be selected from the first stress value in the relaxation late-stage decay sequence, i.e., the stress value at the inflection point, or the initial stress value or the stress value at the end of the sequence. The selection principle for the reference stress value is to ensure that the stress ratio is within a reasonable numerical range and that the logarithmic value is not singular. When the inflection point stress value is selected as the reference stress value, the initial value of the stress ratio sequence is 1, and the corresponding logarithmic stress value is zero. The execution order of the logarithmic domain transformation is as follows: first, traverse each stress value in the relaxation late-stage decay sequence and divide it by the preset reference stress value to obtain the stress ratio sequence; then, perform natural logarithmic or common logarithmic operations on each stress ratio in the stress ratio sequence to generate a logarithmic stress value sequence. Since the stress monotonically decreases and is always positive in the relaxation late stage, the stress ratio is always a positive number less than or equal to 1, and the logarithmic stress value is always a non-positive number. The logarithmic time coordinate carried by the relaxation late-stage decay sequence itself is already a dimensionless value after taking the logarithm of the time parameter, and no repeated transformation is needed. By combining the logarithmic stress value sequence with the corresponding logarithmic time coordinate sequence according to the one-to-one correspondence of the sampling points, a series of two-dimensional data points are formed. The horizontal coordinate of each data point is the logarithmic time coordinate, and the vertical coordinate is the dimensionless logarithmic stress value. All data points together constitute the logarithmic stress-logarithmic time point set.

[0049] Step S253: Perform linear trend fitting on the logarithmic stress-logarithmic time point set, extract the trend line of logarithmic stress changing with logarithmic time, determine the slope of the trend line, and use the slope as the stress decay trend slope. The stress decay trend slope characterizes the rate of decrease of stress on the logarithmic time scale.

[0050] The linear trend fitting process employs the classic least-squares linear regression algorithm. This algorithm uses logarithmic time as the independent variable and logarithmic stress as the dependent variable, searching for a straight line across all discrete points in the logarithmic stress-logarithmic time set that minimizes the sum of squared perpendicular distances from each data point to the line. The solution process for least-squares linear regression includes: calculating the mean of the logarithmic time coordinates of the independent variable and the mean of the logarithmic stress value of the dependent variable; calculating the deviation vectors formed by the differences between each independent variable point and the mean, and the deviation vector formed by the differences between each dependent variable point and the mean; summing the products of the two deviation vectors point by point and dividing by the sum of squared deviations of the independent variable to obtain the slope of the regression line; and subtracting the slope from the mean of the dependent variable and multiplying by the mean of the independent variable to obtain the intercept of the regression line. The regression line, in the logarithmic stress-logarithmic time coordinate system, depicts the average trend of stress decreasing linearly with logarithmic time in the later relaxation stage, and the slope of the line is the slope of the stress decay trend. Since the stress decay of a material over time in the later stages of relaxation follows an approximate power law relationship, which appears as a straight line in a logarithmic coordinate system, the absolute value of the stress decay trend slope directly corresponds to the power law decay exponent. A negative sign of the slope indicates that the stress decreases with increasing time. The stress decay trend slope is a dimensionless value; a larger absolute value indicates a more dramatic decrease in stress with logarithmic time in the later stages of relaxation, and a more significant long-term relaxation process; a smaller absolute value indicates a more gradual decrease in stress, and better long-term stability of the material.

[0051] In this embodiment, least squares linear regression is performed on the logarithmic stress-logarithmic time point set, and the slope of the calculated regression line is a negative number, which is the slope of the stress decay trend.

[0052] Step S254: Based on the power-law decay relationship described by the slope of the stress decay trend, and combined with the physical time step corresponding to the acquisition time step of the single-dimensional decay test sequence, take the initial stress of the relaxation late decay sequence as the benchmark, calculate the remaining ratio of the stress at the end of each physical time step interval relative to the benchmark stress, arrange them in the order of time steps, and form a stress remaining ratio sequence.

[0053] The power-law attenuation relationship described by the stress attenuation trend slope is specifically a function of stress and time, where stress is proportional to the power of the stress attenuation trend slope, and the stress attenuation trend slope is the slope of the straight line extracted in step S253. The physical time step corresponding to the acquisition time step is a fixed physical time interval between the acquisition timestamps of two adjacent response values ​​in the one-dimensional attenuation test sequence. This interval is numerically equal to the reciprocal of the sampling frequency of the data acquisition system in step S100. The execution process for calculating the stress remaining proportion sequence is as follows: First, determine the discrete physical time sequence. Starting from the first time step, the physical time corresponding to the i-th time step is i multiplied by the physical time step. Then, take the initial stress of the relaxation late attenuation sequence as the stress reference value. The physical time corresponding to this initial stress is the physical time at the turning point. Substitute each time value in the physical time sequence into the power-law attenuation relationship formula to calculate the predicted stress value corresponding to that time point. Then, divide the predicted stress value by the stress reference value to obtain the stress remaining proportion of that time step. The stress residual ratio is a dimensionless value between zero and 1, representing the proportion of the stress baseline value remaining at the end of the corresponding physical time step interval. This calculation is performed sequentially for all required time steps, resulting in a set of stress residual ratios arranged in ascending order of time step number, forming a stress residual ratio sequence. The first ratio value in this sequence corresponds to the first time step, and its value is the largest and close to 1. As the time step number increases, the ratio value decreases monotonically, and the rate of decrease is determined by the slope of the stress decay trend.

[0054] Continuing with this embodiment, the physical time step is consistent with the sampling interval of the data acquisition system in step S100. Based on the initial stress of the relaxation late decay sequence, and according to the power law function determined by the stress decay trend slope in step S253, the stress remaining ratio is calculated one by one for each time step, and all ratio values ​​are arranged in order of time step number to form a stress remaining ratio sequence.

[0055] Step S255: Assign the i-th stress residual ratio in the stress residual ratio sequence to the attenuation multiplier of the i-th second diagonal in the Toplitz arrangement structure, where the value of i starts from 1 and increases sequentially to the total number of second diagonals, and all elements on each second diagonal share the attenuation multiplier.

[0056] The total number of secondary diagonals in the Toplitz permutation is equal to the hidden state width minus 1. The secondary diagonals are numbered according to their offsets: offset 1 is the first secondary diagonal, offset 2 is the second, and so on, until the last secondary diagonal with an offset equal to the hidden state width minus 1. The assignment operation assigns the first ratio value in the stress remaining ratio sequence to the first secondary diagonal as its attenuation multiplier, the second ratio value to the second secondary diagonal as its attenuation multiplier, and so on, completing the attenuation multiplier allocation for all secondary diagonals in a one-to-one correspondence. Since the stress remaining ratio sequence itself is a monotonically decreasing sequence with increasing index, and the length of the stress remaining ratio sequence is equal to the total number of secondary diagonals, this assignment operation naturally ensures that secondary diagonals with larger offsets are assigned smaller attenuation multipliers, and secondary diagonals with smaller offsets are assigned larger attenuation multipliers. Regardless of their specific row or column index, all matrix elements on each secondary diagonal use the same attenuation multiplier assigned to that secondary diagonal, as long as the row index minus the column index equals the offset corresponding to that secondary diagonal. This is the meaning of sharing.

[0057] Step S256: Multiply the attenuation multiplier of each secondary diagonal by the value of the main diagonal element to obtain the assignment result of each element on the secondary diagonal. Fill the assignment result into the hidden state cyclic connection weight structure according to the position of the secondary diagonal in the Toplitz arrangement to obtain the secondary diagonal weight arrangement with stress relaxation attenuation characteristics.

[0058] The main diagonal element value is a parameter shared by all elements on the main diagonal with a zero offset in the Toplitz permutation structure. This value is initially assigned during network initialization through random sampling or constant assignment and participates in gradient descent optimization as a trainable parameter during training. For the i-th second diagonal, its decay multiplier is multiplied by the current main diagonal element value, and the product is the unified assignment for all elements on the i-th second diagonal. The filling operation follows the row and column index rules of the Toplitz permutation structure: traverse all rows and columns of the hidden state loop connection weight matrix. For each element in the matrix, calculate the difference between its row index and column index. If the difference is 0, fill in the main diagonal element value; if the difference is i, fill in the product of the decay multiplier of the i-th second diagonal and the main diagonal element value; if the difference is a negative integer k, fill in the corresponding value according to the symmetric design or a separately defined negative offset second diagonal rule. After filling in all the sub-diagonal lines, the hidden state cyclic connection weight matrix presents a complete Toplitz arrangement, and the distribution of element values ​​in the matrix strictly reflects the monotonically decaying law along the offset direction. This sub-diagonal weight arrangement, derived from the power-law decay law of stress relaxation in sensor-sensitive materials, directly solidifies the time decay characteristics of material physical degradation into the connection strength distribution of the cyclic layer's historical state transmission, making the network model naturally conform to the dynamic constraints of physical degradation when extracting temporal features.

[0059] Step S260: Write the initial bias assignment into the forget gate bias term storage unit of the physical constraint loop layer, fill the hidden state loop connection weight storage area of ​​the physical constraint loop layer with the uniform assignment of each diagonal according to the Toplitz arrangement structure, complete the physical constraint configuration of the physical constraint loop layer, and obtain the loop connection weight and bias of the embedded material relaxation time characteristics.

[0060] The forget gate bias storage unit of the physical constraint recurrent layer is a memory space specifically allocated to the forget gate bias scalar in the network parameter data structure, typically stored as a single-precision floating-point scalar. The write operation loads the initial bias assignment determined in step S246 into this storage unit through the parameter assignment interface provided by the deep learning framework, overwriting its default random initialization value. The hidden state cyclic connection weight storage area of ​​the physical constraint recurrent layer is a contiguous memory region allocated to the cyclic connection weight matrix, with a matrix shape equal to the hidden state width multiplied by the hidden state width. The fill operation writes each element to the corresponding memory address in this storage area according to the uniform diagonal assignment obtained in step S256. After both write operations are completed, the two key parameters inside the physical constraint recurrent layer—the forget gate bias and the hidden state cyclic connection weight matrix—both carry physical constraint information derived from the stress relaxation test of the sensitive materials of this batch of sensors. The initial value of the forget gate bias reflects the speed of the material relaxation inflection point, causing the network to be biased towards a memory or forgetting pattern adapted to the material relaxation rate from the very beginning of temporal processing. The Toplitz arrangement of the hidden state cyclic connection weight matrix and the monotonous decay characteristics along the second diagonal reflect the power-law decay law in the later stages of material relaxation, ensuring that the decay pattern of the influence of historical states at different time distances is consistent with the long-term stress decay trend of the material itself. At this point, the physical constraint configuration of the physical constraint cyclic layer is complete, and a cyclic connection weight and bias embedded with the relaxation time characteristics of the sensitive materials of this batch of sensors is ready for subsequent steps to perform physically constrained cyclic feature extraction on the one-dimensional decay test sequence.

[0061] Step S300: Input the one-dimensional decay test sequence into the physical constraint recurrent layer in sequence according to time steps. The physical constraint recurrent layer performs cyclic processing on the input of each time step with physical mechanism constraints by initializing the forget gate bias term and connecting the weights of the Toplitz permutation structure, generating the hidden state vector of each time step, and combining them in time sequence to obtain the physically consistent hidden state sequence.

[0062] In one implementation, step S300 may specifically include the following steps S310 to S360: Step S310: Read the response value of the current time step in the one-dimensional decay test sequence into the input buffer of the physical constraint loop layer as the current input scalar, and at the same time read the previous hidden state vector generated in the previous time step from the state buffer.

[0063] The input buffer is a fast-access storage area allocated at the front end of the physical constraint loop layer to temporarily store input data that will participate in the calculation of the current time step. At the beginning of each time step loop, the sequence read controller retrieves the response value at the corresponding position from the single-dimensional decay test sequence based on the current time step index and loads the response value into the input buffer. The data in the buffer is the current input scalar, and the data type of the current input scalar is a single real number scalar. The state buffer is a dedicated storage area within the physical constraint loop layer for passing hidden states across time steps. After each time step calculation is completed, the newly generated hidden state vector is written to this storage area, and it is read as the source of historical states at the beginning of the next time step calculation. If the current time step is the first time step of the sequence, since there is no hidden state vector generated in the previous time step, the state buffer will return a pre-initialized all-zero vector. The length of this all-zero vector is equal to the preset hidden state width, and each component of the vector is assigned a value of 0. This serves as the previous hidden state vector for the first time step, ensuring the executability of the calculation at the beginning of the sequence.

[0064] Step S320: Perform vector concatenation between the previous hidden state vector and the current input scalar. Connect the components of each dimension of the previous hidden state vector to the current input scalar end to end to obtain a combined vector, which is used as the gated input concatenation vector.

[0065] The input to the vector concatenation process is the previous hidden state vector with a dimension equal to the width of the hidden state and the current input scalar with a dimension of 1. The process first maintains the original order of the components within the previous hidden state vector. Then, the current input scalar is appended immediately after the last component of the vector, becoming the last component of the combined vector. The total dimension of the resulting combined vector equals the hidden state width plus 1; this combined vector is the gated input concatenation vector. Functionally, the gated input concatenation vector uniformly carries the complete content of historical state information and current external observation information. This concatenation vector is fed in one go to the various computational branches of the physical constraint loop layer, including the forget gate weights, input gate weights, and new state generation weights, enabling each gating unit to perform its own linear transformation and gating decisions based on a completely consistent set of information.

[0066] Step S330: Perform a linear transformation on the gated input splicing vector by using the forget gate weights of the physical constraint recurrent layer to obtain the forget gate linear response value. Add the forget gate linear response value to the initial assignment of the forget gate bias term to generate the forget gate comprehensive activation value.

[0067] In one implementation, step S330 may specifically include the following steps S331 to S336: Step S331: Read the weight subset corresponding to the forget gate in the Toplitz permutation structure from the weight storage area of ​​the physical constraint loop layer. The weight subset contains the connection strength values ​​that are weighted for each component of the splicing vector, and the connection strength values ​​also show a decaying arrangement along the hidden state dimension.

[0068] The weight storage area of ​​the physically constrained recurrent layer centrally stores all trainable weight parameters of that layer. The weight subset corresponding to the forget gate is the portion of parameters specifically used for the forget gate's computational branch, cut from the complete weight tensor according to indexing rules. This weight subset contains multiple connection strength values ​​corresponding to each dimension component of the previous hidden state vector and an independent connection strength value corresponding to the current input scalar. The connection strength values ​​corresponding to each dimension component of the previous hidden state vector are also constrained by the Toplitz arrangement structure. Specifically, along the direction of increasing hidden state dimension index, the magnitude of the connection strength values ​​exhibits a gradually decreasing pattern, with components with smaller hidden state dimension indices corresponding to stronger connection strengths and components with larger dimension indices corresponding to weaker connection strengths. This decaying arrangement allows the forget gate to naturally assign different importance weights to different dimension components of the hidden state when weighting the historical state vector, mathematically echoing the decay logic of the Toplitz arrangement structure in the state recurrent connection.

[0069] Step S332: Multiply each component of the gated input concatenation vector by the corresponding connection strength value of the weight subset corresponding to the forget gate point by point to obtain the weighted component set.

[0070] The pointwise multiplication operation strictly follows the principle of one-to-one correspondence between vector components. The first component of the gated input concatenation vector corresponds to the first dimension component of the previous hidden state vector, and is multiplied by the connection strength value of the first dimension of the hidden state in the forget gate weight subset; the second component is multiplied by the connection strength value of the second dimension of the hidden state; and so on until the last component of the gated input concatenation vector, i.e., the current input scalar, is multiplied by the connection strength value of the current input in the forget gate weight subset. All product results are arranged sequentially according to the original order of the multiplicands in the gated input concatenation vector, forming a weighted component set with the same dimensions as the gated input concatenation vector. Each weighted component in this set is the result of selectively scaling the original input component by the forget gate weight coefficients.

[0071] Step S333: Merge all weighted components in the weighted component set, continuously superimpose the values ​​of each weighted component to obtain a scalar output, which serves as the linear response value of the forget gate. The merging process integrates the comprehensive excitation of all dimensions of the spliced ​​vector through the forget gate weights.

[0072] The merging process employs a component-by-component summation method. Specifically, an accumulator variable is initialized to zero. Then, starting from the first weighted component in the weighted component set and proceeding to the last, the value of each weighted component is added to the current value of the accumulator variable, and the result is updated back to the accumulator variable. After the traversal is complete, the final value of the accumulator variable is a single real scalar, which represents the linear response value of the forget gate.

[0073] Step S334: Read the initialization value of the forget gate bias term from the bias storage unit, perform an algebraic sum operation on the initialization value and the forget gate linear response value to obtain the forget gate comprehensive excitation value. The forget gate comprehensive excitation value integrates the combined effects of the current input, historical state and material relaxation bias.

[0074] The bias storage unit is a parameter storage area independent of the weight storage area. It contains only a single scalar parameter: the initialization value of the forget gate bias term after physical constraint assignment. The algebraic sum operation is the addition of the linear response value of the forget gate with the initialization value of the forget gate bias term. The result of the addition is the comprehensive excitation value of the forget gate.

[0075] Step S335: If the total excitation value of the forget gate is greater than the preset upper limit of excitation, it is truncated to the upper limit value; if it is less than the preset lower limit of excitation, it is truncated to the lower limit value, thus obtaining the amplitude-limited excitation value.

[0076] The preset upper and lower bounds of the activation are cutoff boundaries set based on the numerical characteristics of the sigmoid growth curve activation function. When the absolute value of the input activation value is too large, the sigmoid growth curve activation function will enter a saturation region where the gradient is close to zero, causing gradient vanishing during backpropagation. The preset upper bound is set to a positive number that makes the output of the sigmoid growth curve activation function slightly approach the edge of the positive saturation region, and the preset lower bound is set to a negative number that makes the output of the sigmoid growth curve activation function slightly approach the edge of the negative saturation region. The truncation operation involves conditional judgment and assignment replacement: comparing the total activation value of the forget gate with the preset upper bound, if the former is greater than the latter, the total activation value of the forget gate is replaced with the preset upper bound; comparing the total activation value of the forget gate with the preset lower bound, if the former is less than the latter, the total activation value of the forget gate is replaced with the preset lower bound; if neither condition is met, the original value is retained. The output after this truncation process is the amplitude-limited activation value.

[0077] Step S336: Use the amplitude-limiting excitation value as the final forget gate synthesis excitation value used for the activation function to maintain the numerical stability of the gating signal generation process.

[0078] The amplitude-limited excitation value is rewritten into the forget gate comprehensive excitation value variable. The input for the nonlinear activation function to calculate the gated signal in subsequent steps is the stable value after amplitude limiting.

[0079] Step S340: Apply nonlinear activation processing to the forget gate comprehensive excitation value, and map the forget gate comprehensive excitation value to a gating signal value within a continuous opening and closing interval. The gating signal value determines the retention ratio of the previous hidden state vector transmitted to the current time step.

[0080] The nonlinear activation process employs the S-shaped growth curve activation function. Mathematically, this function compresses and maps any input value in the real number domain to a continuous open-closed interval with a lower bound close to zero and an upper bound close to 1. When the total activation value of the forget gate is a large negative number, the output of the S-shaped growth curve activation function approaches zero, indicating that the forget gate tends to almost completely discard historical information. When the total activation value of the forget gate is a large positive number, the output of the S-shaped growth curve activation function approaches 1, indicating that the forget gate tends to almost completely retain historical information. When the total activation value of the forget gate is near zero, the output is approximately the midpoint of the interval, indicating that the forget gate performs a moderate degree of retention of historical information. The mapped gating signal value is a real number located within the continuous open-closed interval. This real number acts as a multiplication factor directly on the multiplication of the previous hidden state vector and the state transit quantity, quantitatively controlling the proportion of the previous hidden state vector information retained in the current time step candidate state.

[0081] Step S350: The current input scalar is weighted by the input weights of the physical constraint loop layer to obtain the input stimulus. The previous hidden state vector is weighted by the hidden state loop connection weights of the Toplitz arrangement structure to obtain the state transfer quantity. The state transfer quantity is multiplied by the gate signal value and then fused with the input stimulus to generate the candidate hidden state.

[0082] The input weights are the set of parameters belonging to the input gate computation branch or the new state generation computation branch in the physical constraint recurrent layer. Their tensor shape is a column vector with a length equal to the hidden state width, and they also include an input bias vector of equal length. The weighted transformation of the input weights on the current input scalar involves multiplying the current input scalar by each weight coefficient in the input weight column vector, and then adding it element-wise to the input bias vector to obtain an input excitation vector with a dimension equal to the hidden state width. This vector carries all the direct contributions of the external observation information at the current time step to the state update. The hidden state recurrent connection weights are the recurrent connection weight matrix with a Toplitz arrangement structure and monotonically decaying second diagonal elements, which has undergone physical constraint configuration in step S260. This matrix performs a standard matrix-vector multiplication linear transformation on the previous hidden state vector, i.e., performing a dot product operation between each row of the recurrent connection weight matrix and the previous hidden state vector to obtain a state transfer vector with a dimension equal to the hidden state width. This state transfer vector represents the information contribution transferred from the historical hidden state to the current time step via the time-distance decaying weighted physical constraint connection. Each dimension component of the state transit vector is multiplied element-wise by the gating signal value generated in step S340. The result of the multiplication is a controlled state transit vector after selective retention or discarding by the forget gate gating signal. The fusion operation performs element-wise real-number addition on the controlled state transit vector and the input stimulus vector according to their corresponding dimension positions. The addition result is the candidate hidden state vector, which integrates information from both historical states and current external inputs.

[0083] Step S360: Apply hyperbolic tangent nonlinear processing to the candidate hidden states to obtain the hidden state vector at the current time step. Store the hidden state vector at the current time step into the state buffer and append it to the hidden state sequence. Recursively push the time steps until all time steps of the one-dimensional decay test sequence have been processed to obtain the physically consistent hidden state sequence.

[0084] The hyperbolic tangent nonlinear processing takes each dimension component of the candidate hidden state vector as an independent input and maps them through the hyperbolic tangent activation function. The hyperbolic tangent function compresses any input in the real domain to a symmetric output interval between -1 and +1. After mapping, a vector with the same dimension as the candidate hidden state vector and all component values ​​between -1 and +1 is obtained. This vector is the hidden state vector at the current time step. The nonlinearity of the hyperbolic tangent activation function gives the hidden state vector the ability to encode complex temporal patterns within a finite amplitude range. At the same time, its symmetric output characteristic about the origin helps maintain the zero-center distribution of the hidden state sequence in the feature space. The newly generated hidden state vector at the current time step is written to the state buffer, overwriting the previously stored hidden state vector, thus updating the contents of the state buffer to prepare for providing historical states for step S310 of the next time step. Simultaneously, the hidden state vector at the current time step is appended to the end of the physically consistent hidden state sequence. The physically consistent hidden state sequence is a dynamically growing ordered list of vectors, initially empty, and expanded by one vector element after each time step processing. The time step index increments sequentially according to the acquisition timestamp, and the complete process from steps S310 to S360 is executed cyclically until all response values ​​in the one-dimensional decay test sequence have been processed in order. At this point, the physically consistent hidden state sequence has accumulated a number of hidden state vectors equal to the number of response values. These vectors are arranged sequentially from the start time step to the end time step of the sequence, forming a complete physically consistent hidden state sequence.

[0085] In one implementation, after step S360, the following steps S370 to S3120 may also be included: Step S370: Extract the hidden state vectors of all generated time steps from the physically consistent hidden state sequence to form the current generated sequence. Perform vector subtraction on the hidden state vectors of adjacent time steps in the current generated sequence to obtain the state drift vector sequence.

[0086] The currently generated sequence is the physically consistent hidden state sequence accumulated and output in step S360, containing all hidden state vectors from the first time step to the current last time step, arranged in ascending order of time steps. The vector subtraction operation operates on each pair of temporally adjacent hidden state vectors in the currently generated sequence. Specifically, the hidden state vector at time step t+1 is used as the subtrahend vector, and the hidden state vector at time step t is used as the subtractor vector. The two vectors of the same dimension are subtracted one-by-one, and the difference vector is the state drift vector corresponding to time step t. The state drift vector describes the instantaneous displacement vector of the hidden state in the feature space when transitioning from one time step to the next. Its magnitude reflects the amplitude of the hidden state change within that time step, and its direction reflects the orientation of the hidden state change in the feature space within that time step. Performing the above vector subtraction on all adjacent time step pairs yields a set of state drift vectors arranged in temporal order, forming a state drift vector sequence.

[0087] Step S380: Calculate the dot product of each state drift vector in the state drift vector sequence with the preset main attenuation direction vector to obtain the drift projection scalar sequence. The preset main attenuation direction vector is determined by the maximum attenuation direction implied by the Toplitz arrangement structure.

[0088] The preset principal decay direction vector is a fixed-unit direction vector with the same dimension as the width of the hidden state. This direction vector is extracted from the Topletz permutation structure of the hidden state cyclic connection weight matrix through mathematical decomposition. The extraction method involves performing eigenvalue decomposition or singular value decomposition on the hidden state cyclic connection weight matrix of the Topletz permutation structure to obtain all eigenvalues ​​or singular values ​​of the matrix. The right eigenvector or right singular vector corresponding to the eigenvalue or maximum singular value with the largest magnitude is selected. After normalizing this vector, the preset principal decay direction vector is obtained. Since the Topletz permutation structure encodes a weight distribution that decays monotonically along the time distance, the direction of the eigenvector corresponding to its maximum eigenvalue naturally points to the principal component direction in which the hidden state decays most slowly and retains information most persistently during cyclic transmission. Geometrically, this direction corresponds to the most important direction of the hidden state decay path. The dot product operation performs an inner product calculation on each state drift vector in the state drift vector sequence with a preset principal decay direction vector. Specifically, it multiplies each dimension component of the state drift vector with the corresponding dimension component of the preset principal decay direction vector, and then sums all the product values ​​to obtain a scalar. Performing the dot product operation on each state drift vector in the sequence yields a set of scalar values ​​arranged in chronological order, forming a drift projection scalar sequence.

[0089] Step S390: Scan the sign change of the drift projection scalar sequence and determine whether there is an alternation of signs from positive to negative or from negative to positive. If the sign remains unchanged throughout the sequence, it is determined that the drift direction has physical monotonicity and consistency, which conforms to the directional attenuation constraint of material stress relaxation.

[0090] The scanning operation begins with the first scalar in the drift projection scalar sequence. The sign of the first scalar is recorded as a reference sign. Then, the second and subsequent scalars in the sequence are read sequentially, and the sign of each scalar is compared with the reference sign. If the signs of all subsequent scalars are identical to the reference sign, the sequence is considered to have unchanged signs. If a scalar sign is found to be opposite to the reference sign, an alternation of signs from positive to negative or from negative to positive is recorded. The directional attenuation constraint of material stress relaxation refers to the fact that during the stress relaxation process of the sensor's sensitive material, the internal stress value evolves along a single, monotonically decreasing direction throughout the entire sequence, without oscillations of repeated increases and decreases in stress value. This physical fact requires that the drift of the hidden state in the feature space must also maintain a single directional attribute along the main attenuation direction; that is, the projected scalar must either always be positive (indicating that the hidden state continuously moves positively along the main attenuation direction) or always be negative (indicating that the hidden state continuously moves negatively along the main attenuation direction) throughout the entire sequence. Therefore, the evolution of the hidden state is considered to be consistent with the physical constraint of directional decay of material stress relaxation if and only if the sign of the drift projection scalar sequence remains unchanged from beginning to end, and the physical monotonicity consistency is verified.

[0091] Step S3100: If the sign of the drift projection scalar sequence remains unchanged, then a physical consistency verification pass flag is added to the physically consistent hidden state sequence, allowing it to proceed to the subsequent stability decay trend analysis process.

[0092] Physical consistency verification is achieved by attaching an identifier, either as a Boolean truth value or a predefined status code, to the metadata field of the physically consistent hidden state sequence. The presence of this identifier indicates that the currently generated physically consistent hidden state sequence has passed the physical consistency verification based on the monotonicity of the main decay direction, and that the drift pattern of its internal hidden state conforms to the basic requirements of the physical law of stress relaxation of sensor sensitive materials. It can be safely used as input data for subsequent stability degradation trend analysis.

[0093] Step S3110: If alternation of signs is detected in the drift projection scalar sequence, it is determined that the physical consistency verification has failed. The value of the forget gate bias term is adjusted by a micro-increment, and the adjusted bias assignment is rewritten to the bias storage unit.

[0094] The alternation of signs indicates that the projection of the hidden state along the main decay direction has reversed direction during time evolution, resulting in oscillating state drift. This oscillation contradicts the physical law that the material stress relaxation process should always maintain directional decay, suggesting that the current physical constraint on the parameters of the loop layer may not adequately constrain the network's drift behavior in the feature space. In particular, the current assignment of the forget gate bias term may cause the gate signal value to be in an unstable range that easily triggers state oscillations. The micro-increment adjustment strategy uses the bias value in the current forget gate bias term storage unit as the base value, and determines the sign and magnitude of the increment based on the detected alternation of signs: if the sign of the drift projection scalar sequence flips from positive to negative, it indicates that the drift direction has reversed in the later stage, and a small negative increment value can be superimposed on the bias base value to slightly reduce the default retention tendency of the forget gate; if the sign flips from negative to positive, a small positive increment value is superimposed on the bias base value to slightly enhance the retention tendency; the adjusted bias assignment is the algebraic sum of the base value and the increment value. The adjusted bias assignment is rewritten to the bias storage unit through the bias initialization interface, overwriting the original bias value.

[0095] Step S3120: Re-execute the loop processing on the one-dimensional decay test sequence with the adjusted bias assignment to generate a new physically consistent hidden state sequence. Repeat the physical consistency verification until it passes, and output the physically consistent hidden state sequence that has passed the verification.

[0096] Using the bias value in the forget gate bias term storage unit updated in step S3110 as the new initial bias value, and keeping all other network parameters unchanged, the complete loop processing flow of steps S310 to S360 is re-executed for the same one-dimensional decay test sequence from the beginning to generate an updated physically consistent hidden state sequence. The physical consistency verification of steps S370 to S390 is then performed again on this new sequence. If symbol alternation is detected again, the bias is fine-tuned and the sequence is regenerated, forming a closed-loop iterative process. This iterative process continues until the drift projection scalar sequence of a generated state drift vector sequence satisfies the condition that the symbols remain unchanged throughout the entire process. The corresponding physically consistent hidden state sequence that has passed the physical consistency verification is then passed as the final output to step S400.

[0097] Step S400: Perform stability decay trend analysis on the physically consistent hidden state sequence, extract the decay path of the hidden state vector in the physically consistent hidden state sequence as time changes, and generate a decay tendency descriptor.

[0098] In one implementation, step S400 may specifically include the following steps S410 to S460: Step S410: Extract all hidden state vectors from the physically consistent hidden state sequence in time step order to form a vector sequence to be analyzed. The dimension of each hidden state vector is consistent with the hidden state width of the physical constraint loop layer.

[0099] In one implementation, step S410 may specifically include the following steps S411 to S416: Step S411: Perform a moving average filter on the hidden state vector in the physically consistent hidden state sequence along the time dimension, and replace the hidden state vector of the center time step with the mean of the hidden state vectors of several adjacent time steps to obtain a smooth hidden state sequence, so as to suppress high-frequency small-amplitude jitter caused by non-physical factors.

[0100] The moving average filtering process operates on the time dimension of the physically consistent hidden state sequence. Its filtering window length is an odd number of time steps to ensure a clear center position. The process begins with the first time step of the physically consistent hidden state sequence. Using the current time step as the center of the filtering window, the hidden state vectors of all time steps within half the window length before and after this center are taken. The arithmetic mean of these vectors is calculated for each corresponding dimensional component. The resulting mean vector is the smoothed hidden state vector for the current center time step. For time steps at the beginning and end of the sequence where the window exceeds the sequence boundary and cannot form a complete window, boundary mirroring or window truncation are used to ensure that each time step outputs a corresponding smoothed hidden state vector. All smoothed hidden state vectors are arranged in the original time step order to form a smoothed hidden state sequence. This sequence effectively filters out high-frequency, small fluctuations in the hidden state that may be caused by data acquisition noise, quantization errors, or environmental perturbations, making the underlying decay trend of the hidden state vector sequence clearer and more discernible.

[0101] Step S412: Obtain the preset stress state projection matrix. The stress state projection matrix is ​​obtained by matrix decomposition of the Toplitz arrangement structure of the physical constraint cycle layer. Its column vectors span the stress-dominant subspace. The dimension of the stress state projection matrix matches the dimension of the hidden state vector.

[0102] The pre-defined stress state projection matrix is ​​generated offline and persistently stored during the model building phase. The generation process involves performing eigenvalue decomposition on the Toplitz arrangement structure of the hidden state cyclic connection weight matrix of the physical constraint loop layer, obtaining all eigenvalues ​​and their corresponding eigenvectors. All eigenvalues ​​are sorted in descending order of magnitude, and a predetermined number of eigenvectors with the largest magnitudes are selected. These eigenvectors are then used as column vectors of the matrix, arranged in descending order of eigenvalues ​​from left to right, forming the stress state projection matrix. The number of rows in the stress state projection matrix equals the width of the hidden state, and the number of columns equals the number of selected pre-defined eigenvectors. These column vectors span a low-dimensional linear subspace in the vector space, which is the stress-dominant subspace. The stress-dominant subspace captures the main eigenmodes contained in the Toplitz weight matrix, representing the dominant component direction in the hidden state cyclic connection where information decays most slowly and stability is most persistent. It effectively distinguishes between components in the hidden state vectors governed by the physical laws of material stress relaxation and components coupled by random perturbations.

[0103] Step S413: Project each smooth hidden state vector in the smooth hidden state sequence onto the stress state projection matrix to obtain the stress state sub-vector at that time step. Subtract the stress state sub-vector from the original smooth hidden state vector to obtain the perturbation sub-vector, thus achieving dimensional decoupling of the hidden state vector.

[0104] The projection operation involves first multiplying the smoothed hidden state vector (considered a row vector) with the stress state projection matrix to obtain a projection coefficient vector. Then, multiplying the projection coefficient vector with the transpose of the stress state projection matrix yields the orthogonal projection vector of the original smoothed hidden state vector onto the stress-dominant subspace. This projection vector is the stress state subvector. The stress state subvector represents the portion of the smoothed hidden state vector governed by the physical laws of stress relaxation. The perturbation subvector is obtained by subtracting the original smoothed hidden state vector from the stress state subvector in a dimension-wise manner. It represents the residual components in the smoothed hidden state vector that cannot be explained by the stress-dominant mode, primarily environmental perturbation coupling components or non-physical noise components. Dimensional decoupling refers to decomposing a hybrid hidden state vector into two physically distinct independent components through the above projection and subtraction operations.

[0105] Step S414: Discard the disturbance sub-vector and retain the stress state sub-vector to obtain the pure stress state vector sequence. The pure stress state vector sequence filters out the environmental disturbance coupling components and retains only the physical components that reflect the stress relaxation inside the material.

[0106] For each time step in the smooth hidden state sequence, only the stress state sub-vector obtained from the projection is retained while the perturbation sub-vector is directly discarded. The stress state sub-vectors retained from all time steps are rearranged in time step order to form a pure stress state vector sequence. This sequence filters out environmental disturbances and non-physical noise to the greatest extent, and the evolution trend of its vectors directly corresponds to the projection trajectory of the pure physical components of stress relaxation inside the sensor's sensitive material in the feature space.

[0107] Step S415: Calculate the covariance structure of the pure stress state vector sequence, and extract the most significant eigendirection of the covariance structure as the principal attenuation direction vector. The most significant eigendirection corresponds to the maximum eigenvalue.

[0108] The calculation process for the covariance structure is as follows: First, calculate the mean of all vectors in the pure stress state vector sequence in each dimension to obtain the mean vector; then, subtract the mean vector from each vector in the sequence to achieve centering; finally, sum the outer products of all centered vectors with themselves and divide by the total number of vectors minus 1 to obtain the covariance matrix. Perform eigenvalue decomposition on this covariance matrix to obtain all eigenvalues ​​and eigenvectors. Select the eigenvalue with the largest value from all eigenvalues; this eigenvalue is the most significant eigenvalue, and its corresponding eigenvector is the principal decay direction vector. Statistically, this direction represents the principal change direction with the largest variance and the richest information retention in the pure stress state vector sequence, and it is also the most important geometric direction of the decay evolution of the hidden state in the feature space.

[0109] Step S416: Perform scalar projection on each stress state vector in the pure stress state vector sequence to the principal attenuation direction vector to obtain a pure one-dimensional projection sequence. The pure one-dimensional projection sequence replaces the one-dimensional projection sequence for subsequent time difference processing.

[0110] The scalar projection operation performs a dot product between each stress state vector and the principal attenuation direction vector in the pure stress state vector sequence. The dot product result is a real scalar, which is the pure one-dimensional projection value for that time step. All pure one-dimensional projection values ​​are arranged in time step order to form a pure one-dimensional projection sequence. In the subsequent time difference processing step S440, this pure one-dimensional projection sequence will replace the original default one-dimensional projection sequence as the difference input.

[0111] Step S420: Calculate the covariance structure of the vector sequence to be analyzed. The covariance structure reflects the relationship between the co-change of each dimension of the hidden state vector over time. Extract the most significant eigendirection of the covariance structure as the principal decay direction vector. The eigenvalue corresponding to the most significant eigendirection is the largest among all eigenvalues.

[0112] In the default analysis path where steps S411 to S416 are not selected, this step directly calculates the covariance structure of the vector sequence to be analyzed formed in step S410. The calculation process is completely consistent with step S415: first, the mean vector of the vector sequence to be analyzed is calculated; then, the mean vector is subtracted from each hidden state vector in the sequence to obtain a centered vector; the covariance matrix is ​​constructed by summing the cross product of the centered vector and itself and taking the average; eigenvalue decomposition is performed on the covariance matrix; and the eigenvector corresponding to the largest eigenvalue is selected as the principal decay direction vector. Statistically, the principal decay direction vector represents the direction with the largest variance in the distribution of hidden state vectors in the vector sequence to be analyzed, and it is also the direction in which the displacement amplitude of the hidden state in the feature space is most significant as it decays over time. Using this direction as the reference direction for scalar projection can preserve and highlight the decay information to the greatest extent.

[0113] Step S430: Project each hidden state vector in the vector sequence to be analyzed onto the main decay direction vector using a scalar method. Calculate the dot product between the hidden state vector and the main decay direction vector to obtain the projection scalar for that time step. The projection scalars of all time steps form a one-dimensional projection sequence in chronological order.

[0114] The scalar projection calculation operates on each hidden state vector in the vector sequence to be analyzed. Each dimensional component of the hidden state vector is multiplied by the corresponding dimensional component of the principal decay direction vector, and all products are summed to obtain a real scalar as the projection scalar for that time step. The same dot product operation is performed on the hidden state vectors for all time steps, and all the resulting projection scalars are arranged in ascending order of time steps, thus forming a one-dimensional projection sequence. This one-dimensional projection sequence compresses the high-dimensional decay trajectory, originally scattered across the width of the hidden states, into a one-dimensional scalar time series along a single principal decay direction, significantly reducing the complexity of subsequent decay trend analysis while retaining the most important components of the decay information.

[0115] Step S440: Perform time difference processing on the one-dimensional projection sequence, subtract the projection scalar of the previous time step from the projection scalar of the next time step to obtain the projection change of adjacent time steps, and form a change sequence from the projection changes of all adjacent time steps.

[0116] Temporal difference processing pairs adjacent projection scalars in the one-dimensional projection sequence. For each pair, a subtraction operation is performed: subtracting the projection scalar at time step t from the projection scalar at time step t+1. The difference is the projection change at the adjacent time step corresponding to time step t. The sign of this change indicates whether the projection scalar increases or decreases within that time step interval, and the absolute value indicates the magnitude of the change. By traversing all adjacent projection scalar pairs in the one-dimensional projection sequence, a set of change values ​​with one less than the length of the one-dimensional projection sequence is obtained. These change values ​​are arranged in time step order to form a change sequence. This change sequence directly describes the time-step rate of change of the hidden state projected along the main decay direction, and is fundamental data for identifying different stages of decay, such as acceleration, deceleration, and stabilization.

[0117] In one implementation, after step S440, the following steps S441 to S446 may also be included: Step S441: Perform sliding window smoothing on the change sequence. Slide a time window of preset length on the change sequence and calculate the arithmetic mean of the change within each window to obtain a local mean change sequence. The local mean change sequence filters out instantaneous fluctuations in the change sequence.

[0118] The sliding window smoothing process selects a fixed odd or even length window. Starting from the first change in the change sequence, the left boundary of the window is aligned with the beginning of the sequence. All change values ​​within the window's coverage area are taken, and their arithmetic mean is calculated. This mean is used as the local mean change value corresponding to the center or end of the window. The window is then slid forward one time step along the time axis, repeating the process of taking changes within the window and calculating the arithmetic mean. During the sliding process, for edge regions at both ends of the sequence that are less than a full window length, partial window truncation or symmetrical filling is used to ensure that each time step has a corresponding local mean change value. All local mean change values ​​are arranged in time step order to form a local mean change sequence. This sequence eliminates instantaneous spikes and glitches caused by noise or random fluctuations in the change sequence, making the macroscopic trend of the rate of change smoother and easier to analyze.

[0119] Step S442: Identify a continuous interval from the local mean change sequence where the value gradually rises from negative to close to zero, and mark the starting position of this continuous interval as the starting point of the decay-gradient transition zone. The starting point of the decay-gradient transition zone indicates that the decay process of the hidden state vector has begun to enter the final stage.

[0120] The identification algorithm scans sequentially from the beginning to the end of the local mean change sequence, comparing each local mean change value with zero. In the early and middle stages of the decay process, the local mean changes are typically persistently negative with relatively large absolute values, indicating a rapid decline in the projected scalar. As the decay process progresses into its later stages, the rate of decline gradually slows, and the absolute values ​​of the local mean changes gradually decrease and approach zero. The scanning algorithm searches the local mean change sequence for a continuous interval exceeding a preset minimum length. Within this interval, while the local mean changes may still be negative, their values ​​exhibit a clear trend of gradually recovering from a relatively negative state and approaching zero. Continuous intervals meeting this condition are identified as decay-slowing transition regions, and the first time step of this interval is marked as the starting point of the decay-slowing transition region. This starting point marks the critical position where the hidden state vector decay process transitions from the main decay stage with a significant rate of decline to the tail segment with an extremely slow rate.

[0121] Step S443: Starting from the beginning of the decay-slowing transition zone, monitor the sign of the change in the subsequent time steps in the change sequence, count the number of consecutive non-negative change steps, and if the number of consecutive steps exceeds the preset number of consecutive steps limit, determine that the decay process of the physically consistent hidden state sequence has entered the stable period.

[0122] The monitoring process uses the sequence of changes after the start of the attenuation slowing transition zone as the monitoring range, reading the projected change value for each adjacent time step. If the currently read change value is greater than or equal to zero, it is determined as a non-negative change, and the continuous non-negative counter is incremented by 1; if the currently read change value is negative, the continuous non-negative counter is reset to zero. Monitoring continues until the count value of the continuous non-negative counter first exceeds the preset continuous step limit. At this point, the current time step of the monitoring point is determined to be the moment when the attenuation process enters the stable period. The preset continuous step limit is a positive integer threshold preset based on long-term stability testing experience of the sensor. Its setting is based on the fact that if the projected change value remains non-negative within a sufficiently long time step window, it indicates that the attenuation of the hidden state along the main attenuation direction has basically stopped, and the sequence has entered a stable stage where no significant system drift occurs.

[0123] Step S444: Extract the hidden state vector corresponding to the first time step of the determination to enter the stable period as the stable reference vector, and at the same time extract the first hidden state vector of the vector sequence to be analyzed as the initial reference vector.

[0124] The first time step determining the entry into the stable period refers to the time step in step S443 where the continuous non-negative counter count value first exceeds the preset duration step limit. The corresponding hidden state vector is extracted from the physically consistent hidden state sequence according to the index of this time step; this is the stable reference vector. The initial reference vector is extracted from the vector sequence to be analyzed or the physically consistent hidden state sequence according to the index of the first time step; this is the hidden state vector at the initial moment of the sequence.

[0125] Step S445: Calculate the vector distance between the stable reference vector and the initial reference vector to obtain the total offset. The vector distance is taken as the magnitude of the difference vector between the two vectors.

[0126] The vector distance calculation first involves subtracting the stable reference vector from the initial reference vector dimension-by-dimensional, resulting in a difference vector with the same dimension as the width of the hidden state. Then, each component of the difference vector is squared, and all squared values ​​are summed. Finally, the square root of the sum is taken, and the resulting non-negative real scalar is the total offset. The total offset quantitatively expresses the total displacement amplitude of the hidden state vector in the feature space during the entire decay process from the start of the sequence to the arrival of stability.

[0127] Step S446: Compare the total offset with the preset offset tolerance. If the total offset exceeds the preset offset tolerance, generate an offset over-limit flag and attach the offset over-limit flag to the decay tendency descriptor so that the decay tendency descriptor carries both decay path direction information and stable period over-limit state information.

[0128] The preset offset tolerance is the maximum allowable total offset value pre-set based on the sensor's long-term stability qualification criteria. This value is determined by comprehensively analyzing statistical data from a large number of historical qualified and unqualified batches, representing the maximum allowable hidden state drift amplitude of the sensor within the test cycle. The comparison operation compares the total offset with the preset offset tolerance. If the total offset is greater than the preset offset tolerance, the offset exceeding the limit flag is set to true; if the total offset is less than or equal to the preset offset tolerance, the offset exceeding the limit flag is set to false. The offset exceeding the limit flag is embedded as supplementary information in the data structure of the degradation tendency descriptor as a Boolean field. This allows the degradation tendency descriptor to carry conclusive information on whether the total offset of the sensor production batch during the stable period exceeds the tolerance limit, in addition to recording the cumulative decay trajectory sequence.

[0129] Step S450: Perform continuous same-sign segment detection on the change sequence, mark the continuous negative change segments as monotonically decaying segments, and record the starting projection scalar and ending projection scalar of each monotonically decaying segment.

[0130] Detection of consecutive segments with the same sign is a one-dimensional sequence segmentation algorithm. It iterates through each value in the sequence of changes, dividing the sequence into several consecutive intervals with the same sign based on the sign of the change value. The algorithm starts with the first change in the sequence, checking if its sign is negative. If negative, a new monotonically decaying segment is started, and the projection scalar corresponding to the previous time step is recorded as the starting projection scalar, which can be extracted from the one-dimensional projection sequence using an index. Subsequent changes are then scanned sequentially. As long as the signs of subsequent changes are consistently negative, these changes belong to the same monotonically decaying segment. When a non-negative change is encountered for the first time, the current monotonically decaying segment ends, and the projection scalar corresponding to the previous change is recorded as the ending projection scalar, which can also be extracted from the one-dimensional projection sequence using an index. The algorithm continues scanning the sequence, starting a new monotonically decaying segment whenever a new negative change is detected, repeating the process of recording the starting and ending projection scalars until the entire sequence has been traversed. Each marked monotonically decaying segment corresponds to a period in which the hidden state continuously and monotonically decreases along the main decay direction, and the projected scalar within the segment is strictly decreasing.

[0131] Step S460: Extract the total decay amount and the segment termination time step of each monotonically decaying segment. Using the segment termination time step as the horizontal axis variable and the total decay accumulated from the beginning to the current segment as the vertical axis variable, generate a cumulative decay trajectory sequence composed of the decay accumulation points of each segment. This sequence serves as a decay tendency descriptor, and the cumulative decay trajectory sequence characterizes the evolution trend of the hidden state decay amplitude as time steps accumulate.

[0132] The total decay of a segment is calculated by subtracting the final projection scalar from the initial projection scalar recorded for each monotonically decaying segment. This difference is always positive, representing the cumulative decrease in the projection scalar along the main decay direction during that monotonically decaying segment. The segment termination time step is the time step index corresponding to the last change in that monotonically decaying segment. The accumulated decay is calculated using an accumulator mechanism. The accumulator is initialized to zero, and the segments are processed sequentially according to their temporal order. When processing the first monotonically decaying segment, the accumulator is added to the total decay of the first segment, and the current accumulator value and the segment termination time step of the first segment are recorded to form the first decay accumulation point. When processing the second monotonically decaying segment, the accumulator continues to accumulate the total decay of the second segment, and the accumulated value and the segment termination time step of the second segment are recorded to form the second decay accumulation point. This process continues until all monotonically decaying segments have been processed. The resulting series of decay accumulation points are arranged in ascending order of segment termination time steps, thus forming the cumulative decay trajectory sequence. This sequence fully depicts the evolution of the hidden state projection scalar as the decay amplitude accumulates over time, starting from the first monotonically decaying segment. The ordinate of the last point in the sequence represents the total cumulative decay amplitude throughout the sequence. If step S446 generates an offset over-limit marker, this marker is encapsulated as an additional field along with the cumulative decay trajectory sequence to form the final decay tendency descriptor.

[0133] Step S500: Using the degradation tendency descriptor as input, perform category mapping processing on the degradation tendency descriptor according to the preset long-term stability judgment rule to obtain the long-term stability identification result corresponding to the sensor production batch.

[0134] In one implementation, step S500 may specifically include the following steps S510 to S560: Step S510: Read the preset long-term stability determination rule base. The rule base contains several stability categories and the corresponding decay boundary contour of each stability category. The decay boundary contour defines the outer periphery of the decay trajectory of the category on the decay amplitude-time span plane.

[0135] The long-term stability determination rule base is a discriminative knowledge base persistently stored on non-volatile storage media in the form of structured files or database tables. The file format can adopt a hierarchical data serialization format. The internal structure of the rule base consists of several rule records. Each rule record is a set of key-value pairs. The key is a stability category identifier, and the value is a set of decay boundary contour data describing the geometric distribution of the decay trajectory of that category. The stability category identifier is a classification label that divides the long-term stability level of the sensor into different levels. For example, it can be divided into several levels such as high stability, medium stability, and low stability categories based on the severity of the sensor's output drift within a specified test period. The decay boundary contour is a set of ordered two-dimensional data points. The horizontal axis of each data point is the time span value, and the vertical axis is the decay amplitude value. These data points are connected sequentially in ascending order of time span value to form one or two outer envelope curves. The area enclosed by the envelope curves defines the distribution range of the typical cumulative decay trajectory belonging to that category on the decay amplitude-time span plane. The upper boundary of this region is the upper limit envelope of the attenuation amplitude that can be reached, and the lower boundary is the lower limit envelope of the attenuation amplitude that can be reached. If the measured cumulative attenuation trajectory falls within or adjacent to the region enclosed by the upper and lower envelopes, it is considered to be highly consistent with this stability category.

[0136] In one implementation, before step S510, the following steps S501 to S506 may also be included: Step S501: Obtain a set of standard sample one-dimensional attenuation test sequences of standard sample batches of sensor sensitive materials under multiple standard aging conditions. Each standard sample sequence in the set of standard sample one-dimensional attenuation test sequences carries a known standard aging condition identifier.

[0137] The standard sample batch is a set of samples specifically selected through statistical sampling from a batch of sensors manufactured under the same raw materials and process conditions as the formal production batch, specifically used to construct the benchmark for judgment rules. Standard aging conditions refer to a series of accelerated aging condition combinations predefined according to sensor product lifespan test standards. Each condition combination consists of key stress parameters such as aging temperature, aging humidity, applied stress amplitude, and aging duration, and each condition combination corresponds to a unique standard aging condition identifier. Under each standard aging condition, the same acquisition process as step S100 is performed on the standard sample sensor to obtain the one-dimensional attenuation test sequence under that condition, and the standard aging condition identifier is permanently bound to this sequence as metadata. All standard sample sequences under all standard aging conditions are summarized to form the standard sample one-dimensional attenuation test sequence set.

[0138] Step S502: For each standard sample one-dimensional decay test sequence, perform the same processing procedure as generating the decay tendency descriptor to generate the corresponding standard sample decay tendency descriptor. The processing procedure includes generating physically consistent hidden state sequences, analyzing stability decay trends, and generating decay tendency descriptors.

[0139] For each standard sample sequence in the set of standard sample one-dimensional decay test sequences, the constructed deep neural network recognition model embedded with the physical degradation mechanism is used to strictly execute the entire process in sequence, including the model parameter configuration in step S200 (forget gate bias initialization and Toplitz weight construction using the same sensor sensitive material relaxation time values ​​as the standard sample batch), the generation of physically consistent hidden state sequences in step S300, and the stability decay trend analysis and decay tendency descriptor generation in step S400, to obtain a standard sample decay tendency descriptor that uniquely corresponds to the standard sample sequence.

[0140] Step S503: Obtain the expected stabilization time corresponding to each standard aging condition, classify each standard sample decay tendency descriptor into different stability categories according to the expected stabilization time, and add a stability category label to each standard sample decay tendency descriptor.

[0141] The expected stabilization time is the rated time expected to maintain the sensor output drift within the allowable limit under each standard aging condition, based on historical reliability test data or industry standards for the sensor product. The classification process involves: pre-setting a set of correspondences between stability categories and expected stabilization time intervals, with each category covering a specific time interval; then reading the standard aging condition identifier for each standard sample sequence and obtaining its corresponding expected stabilization time value through the identifier index; determining the stability category to which the sequence belongs based on the time interval into which the expected stabilization time value falls, and appending the label string corresponding to the stability category to the standard sample degradation tendency descriptor of the sequence as a category label.

[0142] Step S504: Collect the cumulative decay trajectory sequences in all standard decay tendency descriptors under the same stability category, extract the outer envelope curves of these cumulative decay trajectories on the decay magnitude-time span plane, and use the outer envelope curves as the decay boundary contour of the category.

[0143] The aggregation operation groups all standard sample decay tendency descriptors according to stability category labels, extracting all cumulative decay trajectory sequences with the same label to form a set of trajectories of the same category. On the decay amplitude-time span plane, all cumulative decay data points from all sets of trajectories of the same category are plotted on the same coordinate system. The method for extracting the outer envelope curve is to segment along the time span axis with a preset fine step size. At each time segment, the maximum and minimum values ​​of decay amplitude among all trajectories of the same category are found. The maximum decay amplitude values ​​at each time segment are connected to form the upper envelope curve, and the minimum decay amplitude values ​​at each time segment are connected to form the lower envelope curve. The region enclosed by the upper and lower envelope curves is the decay boundary contour of this stability category, which completely depicts the outer periphery of the cumulative decay trajectory distribution of this category on the decay amplitude-time span plane.

[0144] Step S505: Associate and store the decay boundary contours of all stability categories with the corresponding stability category identifiers to obtain the long-term stability determination rule base, and add a version tag to the long-term stability determination rule base, with the version tag corresponding to the standard sample batch information.

[0145] The associated storage serializes the identifier of each stability category and its upper and lower envelope data points together and writes them into the rule base file, and writes a version tag record in the file header or metadata segment. The version tag is a string combination containing information such as the unique batch number of the standard sample, the date the rule base was generated, and the batch identifier of the sensor sensitive material used, which is used to uniquely identify the source of the judgment rule base and its applicable time range.

[0146] Step S506: Add a version tag to the generated long-term stability identification result to trace the version of the judgment rule on which the identification result is based.

[0147] When outputting the long-term stability identification result for the formal production batch in step S500, the version tag string of the long-term stability judgment rule base used this time is filled into the extended field of the result data structure to ensure that each identification conclusion carries the information of its judgment rule base version, so as to achieve full-process identification traceability.

[0148] Step S520: Obtain the cumulative decay trajectory sequence in the decay tendency descriptor. The cumulative decay trajectory sequence consists of cumulative decay total data points arranged in time step order. Each data point contains the termination time step and the cumulative decay total from the start to that time step.

[0149] The cumulative decay trajectory sequence constitutes the main data portion of the decay tendency descriptor generated in step S460. It is directly read from the decay tendency descriptor data structure by data key, yielding an ordered list of two-dimensional data points. Each data point is a key-value pair structure, where the key is the termination time step, and the value is the total cumulative decay accumulated from the start of the sequence to that termination time step. The order of the data points strictly corresponds to the increasing order of the termination time steps.

[0150] Step S530: Perform interpolation smoothing on the cumulative decay trajectory sequence, fill transition points between adjacent data points to obtain a continuous measured decay trajectory curve on the time span axis. The measured decay trajectory curve depicts the cumulative evolution trend of the decay amplitude of the hidden state over time.

[0151] The interpolation smoothing process employs a piecewise cubic spline interpolation algorithm. This algorithm uses data points arranged in ascending order of the final time step in the cumulative decay trajectory sequence as interpolation nodes. For each interval formed by two adjacent nodes, a cubic polynomial curve is constructed based on the function values ​​of these two nodes and the tangent slope values ​​obtained from the overall continuity condition at these two nodes. On each cubic polynomial curve, resampling is performed at a more uniform time step interval than the original data point spacing to generate transition point data within that interval. Connecting all original data points and transition points end-to-end in ascending order of the time span axis forms a mathematically smooth measured decay trajectory curve with continuous first and second derivatives. This curve fully presents the cumulative evolution trend of the hidden state decay amplitude of the sensor production batch over time on the decay amplitude-time span plane.

[0152] Step S540: Sample each decay boundary contour in the rule base at equal time intervals to generate a boundary contour sampling point sequence. At the same time, sample the measured decay trajectory curve at the same density at equal time intervals to obtain a measured sampling point sequence.

[0153] Equal-interval sampling uses a uniform and fixed time coordinate sampling step size, covering the entire analysis interval from the minimum to the maximum value of the time span axis. For the upper and lower envelope curves of each decay boundary contour in the rule base, at each sampling point of the aforementioned uniform time sampling grid, the corresponding upper and lower envelope decay amplitude values ​​are recorded, forming a boundary contour sampling point sequence. Simultaneously, the measured decay trajectory curves are sampled according to the exact same time sampling grid, and the decay amplitude values ​​of the measured decay trajectory curves at each sampling point are recorded, forming a measured sampling point sequence. The measured sampling point sequence and the boundary contour sampling point sequence are strictly aligned one-to-one on the time coordinate, sharing the same set of discrete time coordinate sequences.

[0154] Step S550: For each decay boundary profile, calculate the attenuation amplitude deviation between the measured sampling point sequence and the boundary profile sampling point sequence at the corresponding time position, and merge the attenuation amplitude deviations at all sampling points sequentially along the time axis to obtain the trajectory deviation of the decay boundary profile.

[0155] In one implementation, step S550 may specifically include the following steps S551 to S556: Step S551: Perform time axis alignment processing on the measured sampling point sequence and the boundary contour sampling point sequence to ensure that the sampling points of the two sequences have the same discrete time coordinate sequence.

[0156] The time axis alignment process first compares the time coordinate list of the measured sampling point sequence with the time coordinate list of the boundary contour sampling point sequence to see if they are equal item by item. If there is a slight deviation between the two due to differences in interpolation algorithms or sampling precision, the time coordinate list of the measured sampling point sequence is selected as the reference, and the boundary contour sampling point sequence is resampled using one-dimensional linear interpolation. This ensures that the resampled boundary contour sampling point sequence generates a corresponding boundary amplitude at each reference time coordinate point, thereby ensuring that the two sequences are aligned point by point on the time axis.

[0157] Step S552: For each aligned time coordinate, obtain the attenuation amplitude at the measured sampling point and record it as the measured amplitude. Obtain the attenuation amplitude at the boundary contour sampling point and record it as the contour amplitude. Calculate the absolute value of the difference between the measured amplitude and the contour amplitude, and use it as the point deviation of that time coordinate.

[0158] At each aligned time coordinate, the measured amplitude is directly read from the corresponding element of the measured sampling point sequence. The profile amplitude is selected based on the relative relationship between the measured amplitude and the upper and lower boundaries of the profile: if the measured amplitude is between the lower and upper bound amplitudes, the deviation is not calculated and is directly recorded as zero; if the measured amplitude is higher than the upper bound amplitude, the deviation of the time coordinate is calculated by subtracting the upper bound amplitude from the measured amplitude; if the measured amplitude is lower than the lower bound amplitude, the deviation is calculated by subtracting the measured amplitude from the lower bound amplitude. The value of the deviation is always a non-negative number.

[0159] Step S553: ​​Compose a point deviation sequence from all points aligned to the time coordinates. The length of the point deviation sequence is the same as the number of sampling points.

[0160] Arrange the point deviations calculated at each time coordinate in ascending order to form a point deviation sequence. The number of elements in this sequence is exactly equal to the total number of sampling points.

[0161] Step S554: Calculate the difference between the deviations of adjacent points in the deviation sequence to obtain the deviation change gradient sequence. The deviation change gradient sequence is used to weight subsequent deviation merging.

[0162] The difference calculation operation involves subtracting the deviation of the previous point from the deviation of the next point in the deviation sequence. The resulting differences are arranged in chronological order to form a gradient sequence of deviation changes. This gradient sequence reflects the rate of change of the measured attenuation trajectory's deviation from the contour boundary along the time axis. Regions with increasing absolute gradient values ​​correspond to locations where the deviation increases sharply, and these locations should be given higher weight in the overall evaluation of trajectory deviation.

[0163] Step S555: Using the gradient sequence of deviation changes as the weight sequence, perform weighted merging on the point deviation sequence, multiply the deviation of each point by the weight corresponding to that point and then accumulate along the time axis to obtain the weighted cumulative deviation, and use the weighted cumulative deviation as the trajectory deviation degree.

[0164] The weighted merging process iterates through the deviation sequence while simultaneously traversing the gradient sequence of deviation changes. For each aligned time coordinate point, the deviation value is multiplied by the absolute value or normalized gradient value of the deviation change at that point to obtain the weighted deviation value for that point. The weighted deviation values ​​of all time coordinate points are summed in ascending order along the time axis. The sum is the total weighted cumulative deviation, which is directly assigned as the trajectory deviation of the decay boundary profile.

[0165] Step S556: Use trajectory deviation to measure the overall degree of deviation between the measured decay trajectory and the corresponding decay boundary profile. The smaller the trajectory deviation, the more the measured decay trajectory matches the stability category represented by the profile.

[0166] Trajectory deviation is essentially a comprehensive indicator that measures the degree to which two curves do not overlap on the decay magnitude-time span plane. This indicator integrates information on both the magnitude and duration of deviation. A smaller deviation indicates that the measured decay trajectory falls completely or mostly within the envelope of the decay boundary profile of the stability category, and the decay behavior pattern is highly consistent with that category. A larger deviation indicates that the measured decay trajectory significantly intersects or exceeds the boundary of the category profile, and the decay behavior pattern differs significantly from that category.

[0167] Step S560: Compare the trajectory deviation of all decay boundary profiles, select the stability category corresponding to the decay boundary profile with the smallest trajectory deviation as the long-term stability identification result of the sensor production batch, and output the batch identifier in association with the identification result.

[0168] After calculating the trajectory deviation for each decay boundary profile of all stability categories, a set of deviation values ​​corresponding one-to-one with the stability category identifier is obtained. A minimum search algorithm is used to traverse this set, locating the element with the smallest trajectory deviation value, and obtaining the stability category identifier associated with that element. This identifier represents the long-term stability identification result for the current sensor production batch being evaluated. The unique batch identifier string of the sensor production batch is merged and packaged with the identified stability category label into a single identification result record. Simultaneously, the rule base version flag from step S506 is appended to the record, and the result is output to the production quality traceability database or the batch release decision module of the manufacturing execution system via a data interface, completing the entire automated long-term stability identification process for this batch of sensors.

[0169] Figure 4 A hardware entity diagram of a computer device provided in an embodiment of the present invention, such as... Figure 4 As shown, the hardware entity of the computer device 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

[0170] The memory 1002 stores computer programs that can run on the processor. The memory 1002 is configured to store instructions and applications that can be executed by the processor 1001. It can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) in the processor 1001 and various modules in the computer device 1000. It can be implemented by flash memory or random access memory (RAM).

[0171] When the processor 1001 executes the program, it implements the steps of the sensor production data recognition method using artificial intelligence, as described above. The processor 1001 typically controls the overall operation of the computer device 1000.

[0172] This invention provides a computer storage medium storing one or more programs that can be executed by one or more processors to implement the steps of the sensor production data recognition method using artificial intelligence as described in any of the above embodiments.

[0173] It should be noted that the descriptions of the above storage medium and device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of the present invention, please refer to the descriptions of the method embodiments of the present invention for understanding. The processor described above can be at least one of an Application Specific Integrated Circuit (ASIC), a Digital Signal Processor (DSP), a Digital Signal Processing Device (DSPD), a Programmable Logic Device (PLD), a Field Programmable Gate Array (FPGA), a Central Processing Unit (CPU), a controller, a microcontroller, and a microprocessor. It is understood that the electronic device implementing the above processor function can also be other types, and the embodiments of the present invention do not specifically limit it.

[0174] The aforementioned computer storage media / memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various terminals that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0175] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying sensor production data using artificial intelligence, characterized in that, The method includes: A one-dimensional attenuation test sequence is formed by arranging the response signals obtained through the data acquisition system during the sensor production and testing process in chronological order. The one-dimensional attenuation test sequence consists of multiple response values ​​arranged in chronological order according to the acquisition timestamps, and each response value corresponds to the state of the sensor's sensitive material at a specific time point. A deep neural network recognition model embedding a physical degradation mechanism is constructed. The deep neural network recognition model includes a physical constraint recurrent layer. The forget gate bias term of the physical constraint recurrent layer is initialized based on the relaxation time value of the sensor's sensitive material. The weight structure used to connect the hidden states in the physical constraint recurrent layer is constructed as a Toplitz arrangement structure, and the values ​​of each diagonal element in the Toplitz arrangement structure are set to decrease monotonically with the increase of the propagation time step. The single-dimensional decay test sequence is input into the physical constraint recurrent layer in time steps. The physical constraint recurrent layer performs physical mechanism-constrained recurrent processing on the input of each time step by initializing the forget gate bias term and connecting the weights of the Toplitz permutation structure, generating the hidden state vector of each time step, and combining them in time sequence to obtain the physically consistent hidden state sequence. A stability decay trend analysis is performed on the physically consistent hidden state sequence, and the decay path of the hidden state vector in the physically consistent hidden state sequence as time changes is extracted to generate a decay tendency descriptor. Using the degradation tendency descriptor as input, the degradation tendency descriptor is subjected to category mapping processing according to a preset long-term stability determination rule to obtain the long-term stability identification result corresponding to the sensor production batch.

2. The sensor production data recognition method using artificial intelligence according to claim 1, characterized in that, The deep neural network recognition model embedded with the physical degradation mechanism is constructed as follows: The deep neural network recognition model includes a physically constrained recurrent layer. The forget gate bias term of the physically constrained recurrent layer is initialized based on the relaxation time value of the sensor's sensitive material. The weight structure used to connect the hidden states in the physically constrained recurrent layer is constructed as a Toplitz arrangement, and the values ​​of each diagonal element in the Toplitz arrangement are set to monotonically decay with increasing propagation time step, including: The stress relaxation test record of the sensor-sensitive material is obtained through a constant strain test. The stress relaxation test record contains a stress value sequence arranged in logarithmic time axis order. Each stress value in the stress value sequence corresponds to a sampling time point. The starting stress value of the stress value sequence is the initial stress of the material at the moment when constant strain is applied, and the ending stress value is the residual stress after long-term relaxation. The stress value sequence is subjected to relaxation inflection point location processing. The stress decrease rate is searched in reverse from the end to the beginning of the stress value sequence. When the ratio of the stress decrease rate in adjacent logarithmic time intervals exceeds the preset rate change limit, the boundary point of the time interval is marked as an inflection point, and the logarithmic time coordinate corresponding to the inflection point is extracted as the inflection time value. The logarithmic time coordinate value corresponding to the inflection point is determined as the relaxation time value of the sensor-sensitive material. The relaxation time value characterizes the position of the characteristic inflection point of the material stress relaxation process on the logarithmic time axis and is taken as a dimensionless logarithmic coordinate value as the characteristic time scale of the physical degradation process of the sensor-sensitive material. The forget gate bias term is assigned a value based on the relaxation time value. According to the preset bias determination relationship, the reciprocal of the relaxation time value is mapped to the initial bias assignment value. The bias determination relationship makes the initial bias assignment value and the relaxation time value change in opposite directions. The larger the relaxation time value, the smaller the initial bias assignment value, so as to give the forget gate an activation tendency that is adapted to the relaxation speed of the material at the initial time step. Based on the relaxation time value and the stress decay law over time represented by the stress relaxation test record, the decay multiplier sequence of each diagonal element value in the Toplitz arrangement structure as the transmission time step increases is determined. Each decay multiplier is multiplied by the main diagonal element value to obtain a unified value for all elements on the corresponding diagonal, so that the values ​​of adjacent diagonal elements show a monotonically decreasing trend as the offset increases. The initial bias assignment is written into the forget gate bias term storage unit of the physical constraint loop layer. The uniform assignment of each diagonal is filled into the hidden state loop connection weight storage area of ​​the physical constraint loop layer according to the Toplitz arrangement structure. The physical constraint configuration of the physical constraint loop layer is completed, and the loop connection weight and bias embedded with the relaxation time characteristics of the material are obtained.

3. The sensor production data recognition method using artificial intelligence according to claim 2, characterized in that, Based on the relaxation time value and the stress decay law over time represented by the stress relaxation test record, the decay multiplier sequence of each diagonal element value in the Toplitz arrangement structure is determined as the propagation time step increases. Each decay multiplier is multiplied by the main diagonal element value to obtain a unified value for all elements on the corresponding diagonal, including: The stress value segment from the turning point time value to the end of the sequence is extracted from the stress relaxation test record and used as the relaxation late decay sequence. The relaxation late decay sequence reflects the stress decay process after the material enters the slow relaxation stage. The relaxation late decay sequence is subjected to logarithmic domain transformation. First, each stress value in the sequence is divided by a preset reference stress value to obtain a stress ratio. Then, the logarithm of the stress ratio is taken to generate a dimensionless logarithmic stress value. The time coordinate of the relaxation late decay sequence has been given in the stress relaxation test record according to the logarithmic time axis. The dimensionless logarithmic stress value is combined with the logarithmic time coordinate to form a logarithmic stress-logarithmic time point set. Linear trend fitting is performed on the logarithmic stress-logarithmic time point set to extract the trend line of logarithmic stress changing with logarithmic time. The slope of the trend line is determined and used as the stress decay trend slope. The stress decay trend slope characterizes the rate of decrease of stress on the logarithmic time scale. Based on the power-law attenuation relationship described by the slope of the stress attenuation trend, and combined with the physical time step corresponding to the acquisition time step of the single-dimensional attenuation test sequence, the remaining ratio of stress at the end of each physical time step interval relative to the reference stress is calculated one by one, and arranged in the order of time steps to form a stress remaining ratio sequence. The i-th stress residual ratio in the stress residual ratio sequence is assigned as the attenuation multiplier of the i-th second diagonal in the Toplitz arrangement structure, where the value of i starts from 1 and increases sequentially to the total number of second diagonals, and all elements on each second diagonal share the attenuation multiplier. Multiply the attenuation multiplier of each secondary diagonal by the value of the main diagonal element to obtain the assigned value of each element on that secondary diagonal. Fill the assigned value into the hidden state cyclic connection weight structure according to the position of the secondary diagonal in the Toplitz arrangement to obtain the secondary diagonal weight arrangement with stress relaxation attenuation characteristics.

4. The sensor production data identification method using artificial intelligence according to claim 2, characterized in that, The step of assigning a value to the forget gate bias term based on the relaxation time value, and mapping the reciprocal of the relaxation time value to the initial bias assignment value according to a preset bias determination relationship, includes: The initial stress value and the inflection point stress value are extracted from the stress relaxation test record. The difference between the initial stress value and the inflection point stress value is calculated. The difference is divided by the initial stress value to obtain the stress relaxation amplitude ratio. The stress relaxation amplitude ratio characterizes the degree of stress attenuation of the material when it reaches the inflection point. Using the logarithmic time coordinate value corresponding to the inflection point as the logarithmic time span, the ratio of the stress relaxation amplitude ratio to the logarithmic time span is calculated to obtain the average relaxation rate descriptor, which characterizes the degree of stress decay per unit logarithmic time on the logarithmic time scale. The average relaxation rate descriptor is normalized by taking the maximum and minimum average relaxation rate measurements among similar materials of the sensor's sensitive material as reference limits, and then linearly compressing the average relaxation rate descriptor to the normalized interval to generate a normalized relaxation rate descriptor. The normalized relaxation rate descriptor is input into a preset bias mapping relationship. The bias mapping relationship is constructed such that the larger the normalized relaxation rate descriptor is, the smaller the output bias amount is. When the normalized relaxation rate descriptor is close to the upper bound of the normalization interval, the output bias amount approaches the preset lower bound of the bias. When it is close to the lower bound, the output bias amount approaches the preset upper bound of the bias, thereby obtaining the unlimited initial bias amount. The unlimited initial bias amount is subjected to range limiting processing. If the unlimited initial bias amount exceeds the preset upper limit of the bias tolerance, it is truncated to the upper limit value. If it is lower than the preset lower limit of the bias tolerance, it is truncated to the lower limit value to obtain the limiting bias amount. The limiting bias is used as the initial bias value of the forget gate bias term and stored in the bias initialization interface of the physical constraint loop layer to complete the physical constraint setting of the forget gate bias before the loop process starts.

5. The sensor production data recognition method using artificial intelligence according to claim 1, characterized in that, The process involves sequentially inputting the one-dimensional decay test sequence into the physically constrained recurrent layer at time steps. The physically constrained recurrent layer performs physically mechanism-constrained recurrent processing on the input at each time step using the initialization assignment of the forget gate bias term and the weights of the Toplitz permutation structure, generating a hidden state vector for each time step. This vector is then combined temporally to obtain a physically consistent hidden state sequence, including: The response value of the current time step in the single-dimensional decay test sequence is read into the input buffer of the physical constraint loop layer as the current input scalar, and the previous hidden state vector generated in the previous time step is read from the state buffer. The previous hidden state vector is concatenated with the current input scalar. Each dimension component of the previous hidden state vector is concatenated with the current input scalar to obtain a combined vector, which is used as the gated input concatenation vector. The forgotten gate weights of the physical constraint loop layer are used to perform a linear transformation on the gated input splicing vector to obtain the forgotten gate linear response value. The forgotten gate linear response value is then added to the initial value of the forgotten gate bias term to generate the forgotten gate comprehensive excitation value. A nonlinear activation process is applied to the forget gate comprehensive excitation value, which is then mapped to a gating signal value within a continuous opening and closing interval. The gating signal value determines the retention ratio of the previous hidden state vector transmitted to the current time step. The input scalar is weighted by the input weights of the physical constraint loop layer to obtain the input stimulus. The previous hidden state vector is weighted by the hidden state loop connection weights of the Toplitz arrangement structure to obtain the state transit quantity. The state transit quantity is multiplied by the gate signal value and then fused with the input stimulus to generate a candidate hidden state. Apply hyperbolic tangent nonlinear processing to the candidate hidden state to obtain the hidden state vector at the current time step. Store the hidden state vector at the current time step in the state buffer and append it to the hidden state sequence. Recursively push the time steps until all time steps of the one-dimensional decay test sequence have been processed to obtain the physically consistent hidden state sequence.

6. The sensor production data recognition method using artificial intelligence according to claim 5, characterized in that, The linear transformation of the gated input concatenation vector through the forget gate weights of the physical constraint recurrent layer yields the forget gate linear response value. This forget gate linear response value is then added to the initial assignment of the forget gate bias term to generate the forget gate comprehensive activation value, including: Read the weight subset corresponding to the forget gate in the Toplitz arrangement structure from the weight storage area of ​​the physical constraint loop layer. The weight subset contains the connection strength value that weights each component of the splicing vector, and the connection strength value also shows a decaying arrangement along the hidden state dimension. Each component of the gated input concatenation vector is multiplied point by point with the corresponding connection strength value of the weight subset corresponding to the forget gate to obtain the weighted component set; All weighted components in the weighted component set are merged, and the values ​​of each weighted component are continuously superimposed to obtain a scalar output, which serves as the linear response value of the forget gate. The merging process integrates the comprehensive excitation of all dimensions of the spliced ​​vector through the forget gate weights. The initialization value of the forget gate bias term is read from the bias storage unit, and the value of the initialization value is algebraically summed with the linear response value of the forget gate to obtain the comprehensive excitation value of the forget gate. The comprehensive excitation value of the forget gate integrates the combined effects of the current input, the historical state, and the material relaxation bias. If the total excitation value of the forget gate is greater than the preset upper limit of excitation, it is truncated to the upper limit value; if it is less than the preset lower limit of excitation, it is truncated to the lower limit value, thus obtaining the amplitude-limited excitation value. The amplitude-limiting excitation value is used as the final forget gate synthesis excitation value for the activation function to maintain the numerical stability of the gating signal generation process.

7. The sensor production data recognition method using artificial intelligence according to claim 5, characterized in that, After applying hyperbolic tangent nonlinear processing to the candidate hidden state to obtain the hidden state vector at the current time step, the process further includes: Extract the hidden state vectors of all generated time steps from the physically consistent hidden state sequence to form the current generated sequence. Perform vector subtraction on the hidden state vectors of adjacent time steps in the current generated sequence to obtain the state drift vector sequence. Calculate the dot product of each state drift vector in the state drift vector sequence with the preset main attenuation direction vector to obtain the drift projection scalar sequence. The preset main attenuation direction vector is determined by the maximum attenuation direction implied by the Toplitz arrangement structure. Scan the sign changes of the drift projection scalar sequence to determine whether there is an alternation of signs from positive to negative or from negative to positive. If the sign remains unchanged throughout the sequence, it is determined that the drift direction has physical monotonicity and consistency, which conforms to the directional attenuation constraint of material stress relaxation. If the sign of the drift projection scalar sequence remains unchanged, a physical consistency verification pass flag is added to the physically consistent hidden state sequence, allowing it to proceed to the subsequent stability decay trend analysis process. If alternation of signs is detected in the drift projection scalar sequence, the physical consistency verification is determined to have failed. The value of the forget gate bias term is adjusted by a small increment, and the adjusted bias assignment is rewritten to the bias storage unit. The one-dimensional decay test sequence is reprocessed with the adjusted bias assignment to generate a new physically consistent hidden state sequence. The physical consistency verification is repeated until it passes, and the physically consistent hidden state sequence that passes the verification is output.

8. The sensor production data identification method using artificial intelligence according to claim 1, characterized in that, The step of performing stability decay trend analysis on the physically consistent hidden state sequence, extracting the decay path of the hidden state vector in the physically consistent hidden state sequence over time, and generating a decay tendency descriptor includes: All hidden state vectors are extracted from the physically consistent hidden state sequence in time step order to form a vector sequence to be analyzed. The dimension of each hidden state vector is consistent with the hidden state width of the physical constraint loop layer. Calculate the covariance structure of the vector sequence to be analyzed. The covariance structure reflects the relationship between the co-change of each dimension of the hidden state vector over time. Extract the most significant eigendirection of the covariance structure as the principal decay direction vector. The eigenvalue corresponding to the most significant eigendirection is the largest among all eigenvalues. Each hidden state vector in the vector sequence to be analyzed is scalar projected onto the main decay direction vector, and the dot product of the hidden state vector and the main decay direction vector is calculated to obtain the projection scalar of that time step. The projection scalars of all time steps are arranged in chronological order to form a one-dimensional projection sequence. The one-dimensional projection sequence is subjected to time difference processing. The projection scalar of the previous time step is subtracted from the projection scalar of the next time step to obtain the projection change of adjacent time steps. The projection change of all adjacent time steps constitutes the change sequence. The sequence of changes is subjected to continuous same-sign segment detection. The segments of changes with consecutive negative values ​​are marked as monotonically decaying segments. The starting projection scalar and ending projection scalar of each monotonically decaying segment are recorded. Extract the total decay amount and the segment termination time step of each monotonically decaying segment. Using the segment termination time step as the horizontal axis variable and the total decay accumulated from the beginning to the current segment as the vertical axis variable, generate a cumulative decay trajectory sequence composed of the decay accumulation points of each segment, which serves as the decay tendency descriptor. The cumulative decay trajectory sequence characterizes the evolution trend of the hidden state decay amplitude as time steps accumulate.

9. The sensor production data identification method using artificial intelligence according to claim 1, characterized in that, The process of taking the degradation tendency descriptor as input and performing category mapping processing on the degradation tendency descriptor according to a preset long-term stability determination rule to obtain the long-term stability identification result corresponding to the sensor production batch includes: Read the preset long-term stability determination rule base, which contains several stability categories and the decay boundary contour corresponding to each stability category. The decay boundary contour defines the distribution periphery of the decay trajectory of the category on the decay amplitude-time span plane. Obtain the cumulative decay trajectory sequence in the decay tendency descriptor. The cumulative decay trajectory sequence consists of cumulative decay total data points arranged in time step order. Each data point contains the termination time step and the cumulative decay total from the start to that time step. The cumulative decay trajectory sequence is interpolated and smoothed, and transition points are filled between adjacent data points to obtain a continuous measured decay trajectory curve on the time span axis. The measured decay trajectory curve depicts the cumulative evolution trend of the decay amplitude of the hidden state over time. Each decay boundary contour in the rule base is sampled at equal time intervals to generate a boundary contour sampling point sequence. At the same time, the measured decay trajectory curve is sampled at the same density at equal time intervals to obtain a measured sampling point sequence. For each decay boundary profile, the attenuation amplitude deviation between the measured sampling point sequence and the boundary profile sampling point sequence at the corresponding time position is calculated. The attenuation amplitude deviations at all sampling points are sequentially merged along the time axis to obtain the trajectory deviation of the decay boundary profile. Compare the trajectory deviation of all decay boundary profiles, select the stability category corresponding to the decay boundary profile with the smallest trajectory deviation as the long-term stability identification result of the sensor production batch, and output the batch identifier in association with the identification result.

10. A computer device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 9.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 9.

Images