A machine learning-based medical instrument service prediction system

Abstract

This invention belongs to the field of equipment maintenance prediction technology, specifically relating to a medical device maintenance prediction system based on machine learning. The system includes: a disinfection signal acquisition module for acquiring conductivity time-series signals, pH time-series signals, dialysate temperature time-series signals, real-time output time-series signals of the bicarbonate concentrate proportioning pump, and the nominal concentration of the bicarbonate concentrate; a disinfection signal decoupling module for decoupling temperature compensation and bicarbonate proportioning contribution, obtaining net conductivity deviation time-series signals and net pH deviation time-series signals; a seal degradation identification module for extracting seal degradation feature sequences from the net conductivity deviation time-series signals and net pH deviation time-series signals using machine learning methods, and calculating the equivalent volume index sequence of seal micro-gap; and a maintenance cycle prediction module for fitting the remaining available disinfection cycles and outputting seal maintenance prediction instructions. This invention enables effective maintenance prediction of seals in hemodialysis machines.

Description

Technical Field

[0001] This invention belongs to the field of equipment maintenance prediction technology, specifically relating to a medical device maintenance prediction system based on machine learning. Background Technology

[0002] Hemodialysis machines are essential medical devices for maintaining the lives of patients with kidney failure. Their dialysate circuit system requires regular disinfection during long-term operation to inhibit bacterial growth and ensure the cleanliness and safety of the dialysate. As key components preventing leakage, the seals in the circuit are prone to developing micro-gap dead zones at the sealing contact surfaces after repeated exposure to high temperatures, chemical disinfectant corrosion, and mechanical stress cycles.

[0003] These micro-gaps have complex geometries and small scales, making it difficult for the disinfectant solution to fully replace them during disinfection, resulting in disinfectant residue. When normal dialysate flow is restored after disinfection, the residual disinfectant continues to slowly release ions into the main channel, interfering with the dialysate conductivity. Furthermore, during hemodialysis, the dialysate conductivity itself experiences normal dynamic fluctuations due to multiple factors such as temperature fluctuations and changes in the output of the bicarbonate concentrate pump. These two factors superimpose on the conductivity temporal signal.

[0004] The interference introduced by seal degradation is difficult to separate from normal operating disturbances in terms of signal characteristics. Especially in the early and middle stages of degradation, the dead zone of the micro-gap is small and the ion release is limited. Its impact on conductivity is easily masked by fluctuations in the proportioning pump output and temperature compensation errors. Simply relying on absolute conductivity monitoring will directly misjudge it as a normal operating disturbance. As a result, under complex operating conditions with multiple sources of interference, the sensitivity of identifying the seals of the hemodialysis machine is insufficient, and it is impossible to provide effective predictive maintenance warnings. Summary of the Invention

[0005] (1) Technical problems to be solved The purpose of this invention is to provide a medical device maintenance prediction system based on machine learning to solve the problem that the micro-gap dead zones formed by the degradation of the seals of hemodialysis machines continuously release residual disinfectant ions, which interfere with the conductivity of the dialysate, making it impossible to effectively predict the maintenance of the seals of hemodialysis machines.

[0006] (2) Technical solution To achieve the above objectives, in one aspect, the present invention provides a medical device maintenance prediction system based on machine learning, the system comprising: The disinfection signal acquisition module is used to acquire conductivity time-series signals, pH time-series signals, dialysate temperature time-series signals, and real-time output time-series signals of the bicarbonate concentrate mixing pump after the disinfection operation of the hemodialysis machine is completed, and to obtain the nominal concentration of the bicarbonate concentrate.

[0007] The disinfection signal decoupling module is used to perform temperature compensation on the conductivity time series signal based on the dialysate temperature time series signal to obtain a temperature-compensated conductivity time series signal; and to decouple the temperature-compensated conductivity time series signal and the pH time series signal based on the real-time output time series signal of the bicarbonate concentrate proportioning pump and the nominal concentration of the bicarbonate concentrate to obtain the net conductivity deviation time series signal and the net pH deviation time series signal.

[0008] The sealing component degradation identification module is used to extract the sealing component degradation feature sequence from the time series signals of net conductivity deviation and net pH deviation of each disinfection operation cycle using machine learning methods; and calculates the equivalent volume index of the sealing component micro gap based on the sealing component degradation feature sequence to obtain the equivalent volume index sequence of the sealing component micro gap.

[0009] The maintenance cycle prediction module is used to fit the degradation rate of the equivalent volume index sequence of the micro-gap of the seal to obtain the remaining number of disinfection cycles; when the remaining number of disinfection cycles is lower than the preset service cycle threshold, it outputs the seal maintenance prediction command.

[0010] Furthermore, the disinfection signal decoupling module includes: Obtain the conductivity and pH settings specified in the dialysis prescription from the disinfection operation records.

[0011] The real-time output signal of the bicarbonate concentrate mixing pump was converted with the nominal concentration of the bicarbonate concentrate at each sampling point according to the dilution ratio of the dialysate, so as to obtain the real-time concentration time-series curve of bicarbonate at the dialysate outlet.

[0012] Based on the theoretical relationship between the molar conductivity and concentration of bicarbonate solution at a reference temperature, the real-time concentration time-series curve of bicarbonate at the dialysate outlet is converted into a time-series curve representing the theoretical contribution of bicarbonate to the conductivity of the dialysate. The theoretical contribution time-series curve is then subtracted from the steady-state conductivity contribution corresponding to the nominal concentration of the bicarbonate concentrate to obtain the theoretical contribution deviation time-series curve. Finally, the setpoint conductivity is subtracted from the temperature-compensated conductivity time-series signal, and the theoretical contribution deviation time-series curve is subtracted to obtain the net conductivity deviation time-series signal.

[0013] The real-time concentration time-series curve of bicarbonate at the dialysate outlet is converted into a time-series curve of the theoretical pH contribution of bicarbonate to the dialysate pH. The theoretical pH contribution deviation time-series curve is obtained by subtracting the steady-state pH contribution corresponding to the nominal concentration of bicarbonate concentrate from the pH theoretical contribution time-series curve. The pH setpoint and the pH theoretical contribution deviation time-series curve are then subtracted from the pH time-series signal to obtain the net pH deviation time-series signal.

[0014] Furthermore, the seal degradation identification module includes: The degradation feature extraction submodule is used to map the time series signals of net conductivity deviation and net pH deviation of each disinfection operation cycle onto a two-dimensional phase plane with net conductivity deviation as the horizontal axis and net pH deviation as the vertical axis, so as to obtain the net deviation phase plane trajectory.

[0015] Machine learning methods were used to identify continuous trajectory segments that fall into the fourth quadrant in the net deviation phase plane trajectory, thus obtaining the characteristic trajectory of the phase plane for residual ion release from the seal.

[0016] Within the time interval corresponding to the characteristic trajectory of residual ion release phase plane in the sealing component, the net conductivity deviation at the sampling time corresponding to the maximum distance between each sampling point and the origin of the two-dimensional phase plane is taken as the peak amplitude; the attenuation time constant is obtained by least squares fitting of the time series of net conductivity deviation after the sampling time corresponding to the maximum distance between the sampling point and the origin of the two-dimensional phase plane.

[0017] Based on the overall distribution of the characteristic trajectory of residual ions released from the sealing component in the two-dimensional phase plane, the evolution characteristics of ion composition are extracted.

[0018] The peak amplitude, decay time constant, and ion composition evolution characteristics are used to construct the seal degradation feature vector for the current disinfection operation cycle, and arranged according to the disinfection operation cycle number to form the seal degradation feature sequence.

[0019] Furthermore, the degradation feature extraction submodule includes: The residual ion feature trajectory recognition unit is used to traverse the net deviation phase plane trajectory according to the sampling time sequence, extract continuous time segments with net conductivity deviation greater than zero and net pH deviation less than zero, and obtain candidate fourth quadrant trajectory segments.

[0020] Principal component analysis was performed on the two-dimensional sampling point set composed of the net conductivity deviation and the net pH deviation within each candidate fourth quadrant trajectory segment to obtain the direction vector of the first principal component and the variance contribution rate of the first principal component. The distance sequence from each sampling point within each candidate fourth quadrant trajectory segment to the origin of the two-dimensional phase plane was calculated, and candidate fourth quadrant trajectory segments with a distance sequence that first monotonically increases and then monotonically decreases were extracted to obtain a set of single-peak candidate trajectory segments.

[0021] For each candidate trajectory segment in the set of single-peak candidate trajectory segments, the variance contribution rate of the first principal component and the single-peak morphology index are input into the trained drift event classifier. The candidate trajectory segment with the highest confidence in the classification of drift events is identified as the feature trajectory of the residual ion release phase plane of the sealing component.

[0022] Furthermore, the residual ion feature trajectory recognition unit includes: The net conductivity deviation and net pH deviation at each sampling time within the candidate fourth quadrant trajectory segment are used to form a two-dimensional sampling vector. The product of the two-dimensional sampling vector at each sampling time and its transpose is averaged over all sampling times to obtain a two-dimensional phase plane scatter matrix with the origin of the two-dimensional phase plane as the reference.

[0023] Eigenvalue decomposition is performed on the two-dimensional phase plane scatter matrix to obtain the principal direction eigenvalues, secondary direction eigenvalues, and eigenvectors of the principal direction eigenvalues ​​and secondary direction eigenvalues. The eigenvectors corresponding to the principal direction eigenvalues ​​are taken as the first principal component direction vectors. The variance contribution rate of the first principal component is obtained by dividing the principal direction eigenvalues ​​by the sum of the principal direction eigenvalues ​​and secondary direction eigenvalues.

[0024] Using the fact that the cross element of the net conductivity deviation and the net pH deviation of the two-dimensional phase plane scatter matrix is ​​less than zero, and the conductivity component and pH component of the first principal component direction vector are greater than zero as constraints on the release response of residual ions of acidic disinfectant, the eigenvalue decomposition results of the candidate fourth quadrant trajectory segment are verified, and the first principal component direction vector and the first principal component variance contribution rate that satisfy the release response constraints of residual ions of acidic disinfectant are output.

[0025] Furthermore, the seal degradation identification module includes: The ion composition feature extraction unit is used to divide the absolute value of the pH net deviation component of the first principal component direction vector corresponding to the feature trajectory of the residual ion release phase plane of the seal by the conductivity net deviation component to obtain the ion response ratio of the current disinfection operation cycle; and to calculate the product of the variance contribution rate of the first principal component and the periodic exponential decay value calculated according to the difference of the disinfection operation cycle number to obtain the periodic feature confidence.

[0026] The ion response ratio and the confidence level of the periodic feature are arranged according to the period number to form the ion response ratio confidence level sequence. The dispersion of the ion response ratio is calculated by using the normalized value of the confidence level of the periodic feature as the coefficient for the ion response ratio confidence level sequence, and the historical response ratio dispersion is obtained.

[0027] Based on the confidence sequence and periodic characteristic confidence of the ion response ratio, the monotonic change trend of the ion response ratio with the order of disinfection operation cycles is calculated, and the drift trend of the ion response ratio is obtained.

[0028] The evolution characteristics of ion composition are derived from the historical response ratio dispersion and the ion response ratio drift trend.

[0029] Furthermore, the construction of the drift event classifier includes: Using the peak position of the distance sequence from the sampling point in each candidate fourth quadrant trajectory segment to the origin of the two-dimensional phase plane as the boundary, the monotonically increasing consistency of the distance sequence before the peak position and the monotonically decreasing consistency of the distance sequence after the peak position are calculated, and the product of the two is used to obtain the single-peak morphology index.

[0030] From historical operation records, candidate fourth quadrant trajectory segments extracted during continuous disinfection operation cycles prior to seal replacement events were labeled as positive samples of drift events; candidate fourth quadrant trajectory segments extracted during disinfection operation cycles during the initial service phase after seal replacement events were labeled as negative samples of drift events.

[0031] The positive and negative samples of the drift events are hierarchically sampled to construct a training set; the training set is trained by supervised learning using the variance contribution rate of the first principal component and the unimodal morphology index as classification features to obtain the drift event classifier.

[0032] Furthermore, the seal degradation identification module includes: The equivalent volume calculation submodule is used to calculate the decay time constant of each preceding disinfection operation cycle in the degradation characteristic sequence of the seal by applying an exponential decay weight with the difference in the disinfection operation cycle number as the exponent, and then performing a weighted average calculation to obtain the micro-gap decay time constant of the nth disinfection operation cycle.

[0033] Using the micro-gap decay time constant as the decay parameter, the peak amplitude of the nth period is subjected to multi-preceding period cross-cycle residual correction to obtain the cross-cycle residual corrected peak amplitude.

[0034] The stable operating segment is defined as the time interval from the start of the current disinfection operation cycle sampling window to the start of the characteristic trajectory of the residual ion release phase plane of the seal, excluding all identified candidate fourth quadrant trajectory segments. The mean net conductivity deviation of the stable operating segment is used as the baseline conductivity. The ratio of the absolute value of the mean net pH deviation of the stable operating segment to the mean net conductivity deviation is used as the baseline ion response ratio. The ratio of the square of the baseline ion response ratio to the sum of the square of the baseline ion response ratio and the historical response ratio dispersion is used as the drift trend effective coefficient. The trend-corrected ion response ratio is obtained by summing the product of the ion response ratio drift trend and the drift trend effective coefficient with the baseline ion response ratio. The ion composition-corrected equivalent conductivity is obtained by multiplying the ratio of the baseline ion response ratio to the trend-corrected ion response ratio by the baseline conductivity.

[0035] Peak amplitude of cross-cycle residual correction Micro-gap decay time constant Baseline ion response ratio Baseline conductivity Compared with trend-corrected ion response ratio The equivalent volume index was calculated. The formula for calculating the equivalent volume index is as follows: .

[0036] in, This is the preset disinfectant ion diffusion volume equivalent coefficient.

[0037] Arranged according to the disinfection operation cycle number, a sequence of equivalent volume indexes for the micro-gap of the sealing component is formed.

[0038] Furthermore, the equivalent volume calculation submodule includes: The cumulative cross-cycle time interval from each preceding disinfection operation cycle to the nth disinfection operation cycle is obtained by using the difference between the end time of the disinfection operation in the nth disinfection operation cycle and the end time of the disinfection operation in each preceding disinfection operation cycle.

[0039] Using the cross-cycle residual correction peak amplitude obtained from each preceding disinfection operation cycle as the initial amplitude and the micro-gap decay time constant as the decay parameter, the exponential decay calculation of each preceding disinfection operation cycle according to the corresponding cumulative cross-cycle time interval is performed and then summed to obtain the contribution of the superimposed residual conductivity of multiple preceding cycles.

[0040] The residual corrected peak amplitude across cycles is obtained by subtracting the contribution of residual conductivity from multiple preceding cycles from the peak amplitude of the nth disinfection operation cycle in the seal degradation characteristic sequence.

[0041] Furthermore, the maintenance cycle prediction module includes: The entire disinfection operation cycle sequence of the equivalent volume index of the micro-gap of the sealing component is traversed as candidate degradation inflection points. For each candidate degradation inflection point, the equivalent volume index sequence of the micro-gap of the sealing component is divided into a segment before the inflection point and a segment after the inflection point. Joint regression is performed on the segment before the inflection point and the segment after the inflection point. The two linear fitting lines are made to take the same ordinate value at the candidate degradation inflection point. The candidate degradation inflection point position corresponding to the minimum sum of the squares of the linear fitting residuals of the segment before the inflection point and the segment after the inflection point is taken as the degradation inflection point, thus obtaining the degradation rate before the inflection point and the degradation rate after the inflection point.

[0042] When the degradation rate after the inflection point is greater than the degradation rate before the inflection point, the degradation rate after the inflection point is taken as the current degradation rate; when the degradation rate after the inflection point is not greater than the degradation rate before the inflection point, the degradation rate before the inflection point is taken as the current degradation rate; the remaining number of available disinfection cycles is obtained by dividing the difference between the equivalent volume degradation judgment threshold of the seal and the end value of the equivalent volume index sequence of the micro gap of the seal by the current degradation rate.

[0043] (3) Beneficial effects Compared with the prior art, the beneficial effects of the present invention are: 1. The conductivity time-series signal is decoupled by temperature compensation and bicarbonate ratio contribution through the disinfection signal decoupling module, and the seal degradation identification module extracts the seal degradation features and calculates the equivalent volume index sequence of the seal micro gap using machine learning methods, thereby improving the sensitivity of seal maintenance prediction under multi-source interference conditions.

[0044] 2. The degradation feature extraction submodule maps the time-series signals of net conductivity deviation and net pH deviation of each disinfection operation cycle to a two-dimensional phase plane. Machine learning methods are used to identify the phase plane feature trajectories of residual ion release from seals falling into the fourth quadrant. Based on the constraint of residual ion release response of acid disinfectant, the directionality of seal release events and operating condition disturbances is realized, thereby improving the early sensitivity and anti-background disturbance ability of seal degradation identification. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the module composition of a machine learning-based medical device maintenance prediction system according to Embodiment 1 of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0047] Before providing examples, it is necessary to describe the application scenario of this invention. This invention is applicable to the scenario of online status monitoring and predictive maintenance management of the seals of hemodialysis machines in clinical dialysis centers. After each disinfection operation, the hemodialysis machine enters the dialysate flow recovery phase. During this phase, residual disinfectant ions trapped in the dead zone of the seal's micro-gap are slowly released into the main flow channel, generating weak and limited-duration offset signals on the conductivity and pH sensors. This invention utilizes the sensor signals within the recovery monitoring window after each disinfection operation to separate the characteristic components released by residual ions from the background interference of bicarbonate mixing pump fluctuations and temperature disturbances. By tracking the cycle-by-cycle evolution of the equivalent volume index of the seal's micro-gap, the invention predicts the timing of seal maintenance.

[0048] Example 1: As Figure 1 As shown, this embodiment provides a medical device maintenance prediction system based on machine learning, the system comprising: The disinfection signal acquisition module is used to acquire conductivity time-series signals, pH time-series signals, dialysate temperature time-series signals, and real-time output time-series signals of the bicarbonate concentrate mixing pump after the disinfection operation of the hemodialysis machine is completed, and to obtain the nominal concentration of the bicarbonate concentrate.

[0049] The disinfection signal decoupling module is used to perform temperature compensation on the conductivity time series signal based on the dialysate temperature time series signal to obtain a temperature-compensated conductivity time series signal; and to decouple the temperature-compensated conductivity time series signal and the pH time series signal based on the real-time output time series signal of the bicarbonate concentrate proportioning pump and the nominal concentration of the bicarbonate concentrate to obtain the net conductivity deviation time series signal and the net pH deviation time series signal.

[0050] The sealing component degradation identification module is used to extract the sealing component degradation feature sequence from the time series signals of net conductivity deviation and net pH deviation of each disinfection operation cycle using machine learning methods; and calculates the equivalent volume index of the sealing component micro gap based on the sealing component degradation feature sequence to obtain the equivalent volume index sequence of the sealing component micro gap.

[0051] The maintenance cycle prediction module is used to fit the degradation rate of the equivalent volume index sequence of the micro-gap of the seal to obtain the remaining number of disinfection cycles; when the remaining number of disinfection cycles is lower than the preset service cycle threshold, it outputs the seal maintenance prediction command.

[0052] For example, the disinfection operation ends when all steps of the hemodialysis machine disinfection procedure (including disinfectant circulation and subsequent rinsing) are completed and the dialysate circuit switches to normal flow mode. After the disinfection operation, a 20-minute monitoring window is activated to simultaneously collect conductivity and pH time-series signals at the dialysate outlet at a sampling rate of 2Hz, collecting a total of 2400 sampling points. The dialysate temperature time-series signal is simultaneously collected at a sampling rate of 2Hz by a temperature sensor installed at the dialysate outlet. The real-time output time-series signal of the bicarbonate concentrate mixing pump is synchronously output by the hemodialysis machine control system at a frequency of 2Hz through a digital communication interface, aligning with the sampling times of the conductivity and pH signals. The nominal concentration of the bicarbonate concentrate is a fixed parameter given by the dialysis prescription.

[0053] The conductivity of dialysate is easily affected by temperature. If the temperature effect is not compensated for, the offset caused by temperature drift will overwhelm the release signal of the seal. The conductivity temperature coefficient is used to correct each sampling point of the conductivity time series signal to the equivalent conductivity at the reference temperature by multiplying the difference between the temperature at each sampling point in the dialysate temperature time series signal and the reference temperature of 25°C. This yields the temperature-compensated conductivity time series signal. The conductivity temperature coefficient is determined based on the conductivity-temperature characteristics of the dialysate near the reference temperature of 25°C.

[0054] Based on the real-time output time-series signal of the bicarbonate concentrate mixing pump and the nominal concentration of the bicarbonate concentrate, the bicarbonate ratio contribution of the temperature-compensated conductivity time-series signal and the pH time-series signal are decoupled to obtain the net conductivity deviation time-series signal and the net pH deviation time-series signal.

[0055] Using the decoupled net conductivity deviation and net pH deviation as inputs, the process enters the degradation feature extraction submodule. This submodule identifies the residual ion release phase plane feature trajectory of the seal within the net deviation phase plane trajectory. In this disinfection operation cycle, trajectory segment P2 is confirmed as the residual ion release phase plane feature trajectory of the seal. The submodule extracts the peak amplitude, decay time constant, and ion composition evolution characteristics to construct the seal degradation feature vector for this disinfection operation cycle. This vector, along with the degradation feature vectors from previous disinfection operation cycles, is arranged according to the cycle number to form the seal degradation feature sequence.

[0056] The equivalent volume calculation submodule takes the peak amplitude and decay time constant sequence in the seal degradation characteristic sequence as input, and calculates the equivalent volume index of this disinfection operation cycle through multi-preceding cycle cross-cycle residual correction and heterogeneous diffusion equivalent volume calculation. The equivalent volume indexes of the first 8 disinfection operation cycles are 0.0262mL, 0.0269mL, 0.0271mL, 0.0275mL, 0.0280mL, 0.0283mL, 0.0286mL, and 0.0289mL, respectively, and are arranged according to the disinfection operation cycle number to form the equivalent volume index sequence of the seal micro gap.

[0057] The preset usage period threshold is a configurable parameter of the system. Operators enter this threshold into the system configuration interface based on actual operating conditions, and it remains fixed until the next manual modification. The configuration value is determined by the comprehensive sealing component procurement cycle, the frequency of hemodialysis machine disinfection, and safety margin: based on the sealing component procurement cycle... Weekly disinfection Next, the safety margin is Taking a week as an example, using a periodic threshold The recommended configuration formula is: Taking the specific parameters of this embodiment as an example: the procurement cycle for seals is 2 weeks, disinfection is performed 6 times per week, and the safety margin is 1 week. Therefore, the recommended configuration value for the cycle threshold is 18 times. After the operator writes 18 into the system configuration item for the cycle threshold, it will serve as a fixed judgment benchmark in each subsequent maintenance prediction.

[0058] The current remaining number of available disinfection cycles is 44, which is greater than 18, so no seal maintenance prediction command will be output for now. When the remaining number of available disinfection cycles obtained by fitting the subsequent degradation rate is less than 18, a seal maintenance prediction command will be output to prompt maintenance personnel to replace the seals before the current stock is exhausted.

[0059] The disinfection signal decoupling module includes: Obtain the conductivity and pH settings specified in the dialysis prescription from the disinfection operation records.

[0060] The real-time output signal of the bicarbonate concentrate mixing pump was converted with the nominal concentration of the bicarbonate concentrate at each sampling point according to the dilution ratio of the dialysate, so as to obtain the real-time concentration time-series curve of bicarbonate at the dialysate outlet.

[0061] Based on the theoretical relationship between the molar conductivity and concentration of bicarbonate solution at a reference temperature, the real-time concentration time-series curve of bicarbonate at the dialysate outlet is converted into a time-series curve representing the theoretical contribution of bicarbonate to the conductivity of the dialysate. The theoretical contribution time-series curve is then subtracted from the steady-state conductivity contribution corresponding to the nominal concentration of the bicarbonate concentrate to obtain the theoretical contribution deviation time-series curve. Finally, the setpoint conductivity is subtracted from the temperature-compensated conductivity time-series signal, and the theoretical contribution deviation time-series curve is subtracted to obtain the net conductivity deviation time-series signal.

[0062] The real-time concentration time-series curve of bicarbonate at the dialysate outlet is converted into a time-series curve of the theoretical pH contribution of bicarbonate to the dialysate pH. The theoretical pH contribution deviation time-series curve is obtained by subtracting the steady-state pH contribution corresponding to the nominal concentration of bicarbonate concentrate from the pH theoretical contribution time-series curve. The pH setpoint and the pH theoretical contribution deviation time-series curve are then subtracted from the pH time-series signal to obtain the net pH deviation time-series signal.

[0063] For example, the conductivity and pH settings specified in the dialysis prescription remain fixed within the same sterilization cycle and are read directly from the sterilization operation record. Taking a set of prescription parameters as an example, the conductivity setting is 14.0 mS / cm, and the pH setting is 7.35.

[0064] The unit of the real-time output flow rate of the bicarbonate concentrate mixing pump is mL / min. Taking a dilution ratio of 1:34 for solution B as an example, the total volume of dialysate consists of 1 part solution B and 34 parts treated water, with the volume fraction of solution B in the dialysate being 1 / 35. The nominal concentration of the bicarbonate concentrate is 700 mmol / L. When the mixing pump is operating at its rated flow rate, the bicarbonate concentration at the dialysate outlet is 700 × 1 / 35 = 20.0 mmol / L. For conversion at each sampling point, the ratio of the real-time output flow rate of the mixing pump at the current sampling time to the rated output flow rate is multiplied by 20.0 mmol / L to obtain the real-time bicarbonate concentration curve at the dialysate outlet, in mmol / L.

[0065] The conversion of the theoretical contribution of conductivity is based on Kohlrausch's law of independent ion migration. Within the actual electrolyte concentration range of the dialysate (bicarbonate approximately 20 mmol / L, within the applicable range for dilute solutions), the change in dialysate conductivity caused by a change in bicarbonate concentration is equal to the product of the sum of the molar conductivities of each ion that change synchronously with the bicarbonate ratio and the change in concentration. Solution B's main component is sodium bicarbonate. The molar conductivity of bicarbonate ions at a reference temperature of 25℃ is approximately 44.5 mS·cm² / mmol, and the accompanying molar conductivity of sodium ions is approximately 50.1 mS·cm² / mmol, with a sum of approximately 94.6 mS·cm² / mmol. The reference temperature of 25℃ is consistent with the reference temperature of the temperature compensation step, ensuring that the decoupled calculation of the conductivity channel is performed under the same temperature reference. The values ​​of each sampling point on the time-series curve of the real-time bicarbonate concentration at the dialysate outlet are multiplied by 94.6 mS·cm² / mmol and converted to units to obtain the time-series curve of the theoretical contribution of conductivity. Substituting the nominal concentration of the dialysate outlet (20.0 mmol / L) corresponding to the nominal concentration of the bicarbonate concentrate into the same relationship, the steady-state conductivity contribution is obtained as 20.0 × 0.0946 = 1.89 mS / cm. Subtracting the steady-state conductivity contribution of 1.89 mS / cm from the theoretical conductivity contribution time-series curve yields the theoretical conductivity contribution deviation time-series curve. When the mixing pump is running at its rated flow rate, the values ​​of the theoretical conductivity contribution deviation time-series curve are close to zero at all points, with a non-zero deviation only appearing when the mixing pump output deviates from the rated value. Subtracting the conductivity setpoint of 14.0 mS / cm from the temperature-compensated conductivity time-series signal, and subtracting the theoretical conductivity contribution deviation time-series curve, yields the net conductivity deviation time-series signal.

[0066] The pH conversion is based on the Henderson-Hasselbalch equation: pH = pKa1 + log([HCO3⁻] / (α·pCO2)), where pKa1 is taken as the first dissociation constant of carbonic acid at 37℃ (6.10), α is the solubility coefficient of CO2 in dialysate at 37℃ (approximately 0.0306 mmol / (L·mmHg)), and pCO2 is the partial pressure of CO2 in dialysate. During the re-intervention phase, the ratio of solution A (acid concentrate) is maintained near the prescription setting by servo control, and pCO2 is approximately constant at 40 mmHg (pCO2 depends on the specific hemodialysis machine circuit structure and sensor position, and should be determined by pH steady-state calibration during system initialization according to the actual machine model). Therefore, α·pCO2 = 0.0306 × 40 ≈ 1.22 mmol / L. Substituting the values ​​of each sampling point of the real-time bicarbonate concentration time series curve at the dialysate outlet into the Henderson-Hasselbalch equation, the theoretical contribution time series curve of pH is obtained. Substituting the nominal concentration of bicarbonate concentrate (20.0 mmol / L), the steady-state pH contribution is calculated as 6.10 + log(20.0 / 1.22) = 6.10 + 1.21 = 7.31. Subtracting the steady-state pH contribution of 7.31 from the theoretical pH contribution time-series curve yields the theoretical pH contribution deviation time-series curve. Subtracting the pH setpoint of 7.35 from the pH time-series signal, and then subtracting the theoretical pH contribution deviation time-series curve, yields the net pH deviation time-series signal.

[0067] It should be noted that the difference between the steady-state theoretical contribution of conductivity (1.89 mS / cm) and the set conductivity value (14.0 mS / cm), and the difference between the steady-state theoretical contribution of pH (7.31) and the set pH value (7.35) reflect the combined contribution of other electrolyte components such as sodium, potassium, calcium, magnesium, and chloride in the dialysate to each indicator, and are unrelated to the bicarbonate ratio.

[0068] The seal degradation identification module includes: The degradation feature extraction submodule is used to map the time series signals of net conductivity deviation and net pH deviation of each disinfection operation cycle onto a two-dimensional phase plane with net conductivity deviation as the horizontal axis and net pH deviation as the vertical axis, so as to obtain the net deviation phase plane trajectory.

[0069] Machine learning methods were used to identify continuous trajectory segments that fall into the fourth quadrant in the net deviation phase plane trajectory, thus obtaining the characteristic trajectory of the phase plane for residual ion release from the seal.

[0070] Within the time interval corresponding to the characteristic trajectory of residual ion release phase plane in the sealing component, the net conductivity deviation at the sampling time corresponding to the maximum distance between each sampling point and the origin of the two-dimensional phase plane is taken as the peak amplitude; the attenuation time constant is obtained by least squares fitting of the time series of net conductivity deviation after the sampling time corresponding to the maximum distance between the sampling point and the origin of the two-dimensional phase plane.

[0071] Based on the overall distribution of the characteristic trajectory of residual ions released from the sealing component in the two-dimensional phase plane, the evolution characteristics of ion composition are extracted.

[0072] The peak amplitude, decay time constant, and ion composition evolution characteristics are used to construct the seal degradation feature vector for the current disinfection operation cycle, and arranged according to the disinfection operation cycle number to form the seal degradation feature sequence.

[0073] For example, the disturbance from the bicarbonate mixing pump causes the net conductivity deviation and the net pH deviation to change in the same direction, and their trajectories fall into the first or third quadrant in the two-dimensional phase plane; the release of residual ions from the acidic disinfectant causes the net conductivity deviation to be positive and the net pH deviation to be negative, and their trajectories fall into the fourth quadrant in the phase plane. After mapping to the phase plane, the quadrant assignment of the trajectory provides a directional constraint independent of amplitude, making the two types of disturbances distinguishable even in the early stages of degradation when the peak value of the net conductivity deviation is only about 0.10 to 0.15 mS / cm. In this disinfection operation cycle, the degradation feature extraction submodule confirmed that trajectory segment P2 is a phase plane characteristic trajectory of residual ion release from the seal.

[0074] Within the time interval corresponding to the characteristic trajectory of the residual ion release phase plane of the sealing component, the net conductivity deviation at the sampling time when the distance between each sampling point and the origin of the two-dimensional phase plane is the maximum is taken as the peak amplitude; the least squares fitting is performed on the time series of net conductivity deviation after the sampling time to obtain the decay time constant.

[0075] In this disinfection cycle, the peak of the distance sequence in segment P2 occurred at the 40th sampling point within P2, corresponding to approximately 230 seconds after the disinfection operation ended. The peak amplitude of this disinfection cycle was defined as the net conductivity deviation of 0.152 mS / cm at the time of sampling. After the peak, the residual ion concentration continued to decrease. Although the net conductivity deviation had left the fourth quadrant after the end of segment P2 (255 seconds), the decreasing trend continued. For the time series of net conductivity deviations from the peak to the end of this sampling (a total of 1940 sampling points), using the peak amplitude of 0.152 mS / cm as the initial value and the decay time constant as the parameter to be estimated, the sum of the squares of the differences between the observed values ​​and the calculated exponential decay values ​​at each sampling point was minimized, and a least-squares fit was performed. The resulting decay time constant for this disinfection cycle was 337 s.

[0076] The evolution characteristics of ion composition are extracted periodically from the overall distribution direction of the characteristic trajectory of residual ions released from the sealing element in the two-dimensional phase plane. Then, the distribution dispersion and monotonic change trend of the ion response ratio are calculated for the cross-period directional characteristic sequence. The peak amplitude, decay time constant, and ion composition evolution characteristics constitute the sealing element degradation characteristic vector for this disinfection operation cycle, which is arranged sequentially according to the disinfection operation cycle number to form the sealing element degradation characteristic sequence.

[0077] The degradation feature extraction submodule includes: The residual ion feature trajectory recognition unit is used to traverse the net deviation phase plane trajectory according to the sampling time sequence, extract continuous time segments with net conductivity deviation greater than zero and net pH deviation less than zero, and obtain candidate fourth quadrant trajectory segments.

[0078] Principal component analysis was performed on the two-dimensional sampling point set composed of the net conductivity deviation and the net pH deviation within each candidate fourth quadrant trajectory segment to obtain the direction vector of the first principal component and the variance contribution rate of the first principal component. The distance sequence from each sampling point within each candidate fourth quadrant trajectory segment to the origin of the two-dimensional phase plane was calculated, and candidate fourth quadrant trajectory segments with a distance sequence that first monotonically increases and then monotonically decreases were extracted to obtain a set of single-peak candidate trajectory segments.

[0079] For each candidate trajectory segment in the set of single-peak candidate trajectory segments, the variance contribution rate of the first principal component and the single-peak morphology index are input into the trained drift event classifier. The candidate trajectory segment with the highest confidence in the classification of drift events is identified as the feature trajectory of the residual ion release phase plane of the sealing component.

[0080] For example, 2400 sampling points are mapped onto a two-dimensional phase plane with net conductivity deviation as the horizontal axis and net pH deviation as the vertical axis to form the net deviation phase plane trajectory of this disinfection operation cycle.

[0081] Traversing the phase plane trajectory according to the sampling time sequence, for each sampling point, it was determined whether the net conductivity deviation was greater than zero and the net pH deviation was less than zero. Segments of sampling points that continuously met the conditions were divided into candidate fourth quadrant trajectory segments. A total of 3 segments were identified within this monitoring window. Trajectory segment P1 started at 82 seconds after the disinfection operation ended and lasted for 18 seconds (36 sampling points); trajectory segment P2 started at 210 seconds and lasted for 45 seconds (90 sampling points); and trajectory segment P3 started at 680 seconds and lasted for 9 seconds (18 sampling points). The simultaneous occurrence of net conductivity deviation and net pH deviation falling into the fourth quadrant is a necessary condition, but not a sufficient condition, for the release of residual ions from the seal. The servo control circuit of the hemodialysis machine continuously adjusts the bicarbonate proportioning pump in the early stage of dialysate flow recovery. Excessive adjustment can also cause a short-term increase in conductivity and a decrease in pH. These disturbance segments have weak spatial extension in the phase plane and are different from the trajectory characteristics formed by the continuous diffusion of residual ions from the seal.

[0082] Principal component analysis was performed on the two-dimensional sampling point sets of trajectory segments P1, P2, and P3 to obtain the direction vector of the first principal component and the variance contribution rate of the first principal component for each segment.

[0083] The Euclidean distance from each sampling point within each candidate trajectory segment to the origin of the two-dimensional phase plane was calculated, forming a distance sequence for each segment. The origin of the two-dimensional phase plane, rather than the centroid of the trajectory segment, was chosen as the reference point because the net deviations of conductivity and pH approach zero under undisturbed conditions, and the origin represents the normal operating baseline state of the dialysate circuit. When residual ions from the seal diffuse from the micro-gap into the main channel, the net deviation starts from zero, reaching its maximum distance from the origin at the peak of ion release flux, and then decreases as ions are transported and diluted. This physical process gives the distance sequence of the actual release event a single-peak shape that first monotonically increases and then monotonically decreases. The distance sequence of P1 increases from 0.06 to 0.13 in the first 18 sampling points and decreases from 0.13 to 0.07 in the last 18 sampling points. The distance sequence of P2 increases from 0.09 to 0.22 in the first 40 sampling points and decreases from 0.22 to 0.08 in the last 50 sampling points. Both satisfy the condition of increasing first and then decreasing, and are included in the single-peak candidate trajectory segment set. The distance sequence of P3 continuously decreases monotonically in 18 sampling points and has no peak structure, so it is excluded.

[0084] Both P1 and P2 passed the binary morphology screening, but binary screening itself cannot distinguish between real seal release events and servo disturbance segments with poor single-peak quality, nor can it utilize the trajectory linear extension information extracted by principal component analysis. Directly setting a hard threshold for the first principal component variance contribution rate has poor adaptability across different degradation stages and different machine individuals. Therefore, the first principal component variance contribution rate and the single-peak morphology index are used as continuous features input to the drift event classifier, which makes a comprehensive judgment. In this embodiment, the single-peak morphology index of P1 is 0.68, and that of P2 is 0.93. After inputting into the drift event classifier, the classification confidence of drift events for P1 is 0.38, and for P2 it is 0.87. P2 is confirmed as the seal residual ion release phase plane characteristic trajectory of this disinfection operation cycle.

[0085] The residual ion feature trajectory recognition unit includes: The net conductivity deviation and net pH deviation at each sampling time within the candidate fourth quadrant trajectory segment are used to form a two-dimensional sampling vector. The product of the two-dimensional sampling vector at each sampling time and its transpose is averaged over all sampling times to obtain a two-dimensional phase plane scatter matrix with the origin of the two-dimensional phase plane as the reference.

[0086] Eigenvalue decomposition is performed on the two-dimensional phase plane scatter matrix to obtain the principal direction eigenvalues, secondary direction eigenvalues, and eigenvectors of the principal direction eigenvalues ​​and secondary direction eigenvalues. The eigenvectors corresponding to the principal direction eigenvalues ​​are taken as the first principal component direction vectors. The variance contribution rate of the first principal component is obtained by dividing the principal direction eigenvalues ​​by the sum of the principal direction eigenvalues ​​and secondary direction eigenvalues.

[0087] Using the fact that the cross element of the net conductivity deviation and the net pH deviation of the two-dimensional phase plane scatter matrix is ​​less than zero, and the conductivity component and pH component of the first principal component direction vector are greater than zero as constraints on the release response of residual ions of acidic disinfectant, the eigenvalue decomposition results of the candidate fourth quadrant trajectory segment are verified, and the first principal component direction vector and the first principal component variance contribution rate that satisfy the release response constraints of residual ions of acidic disinfectant are output.

[0088] For example, taking trajectory segment P2 (90 sampling points) identified in this disinfection operation cycle as an example, the net conductivity deviation of each sampling point within P2 is between 0.02 and 0.15 mS / cm, and the net pH deviation is between -0.21 and -0.03, all falling within the fourth quadrant. Let the net conductivity deviation at the i-th sampling time be denoted as... The net pH deviation is recorded as This forms a two-dimensional sampling vector. The mean of the product of the two-dimensional sampling vector and its transpose at each sampling time is calculated to obtain the two-dimensional phase plane scatter matrix with the origin of the two-dimensional phase plane as the reference. The four elements are: =0.0098, = =−0.0078, =0.0102.

[0089] The scatter matrix is ​​defined as the mean of the cross product of the coordinate vectors at each sampling time, rather than the mean-removed covariance matrix. Both the net conductivity deviation and net pH deviation are based on the steady-state theoretical values ​​under normal operation, approaching zero under undisturbed conditions. The origin represents the normal operating baseline state of the circuit. If a mean-removed covariance matrix is ​​used, the calculation baseline will shift to the centroid of the trajectory segment, treating the overall amplitude of the entire trajectory relative to the origin in the phase plane as zero. Consequently, the overall amplitude information reflected by the continuous release of residual ions from the seal is lost.

[0090] For the scatter matrix Perform eigenvalue decomposition, principal direction eigenvalues =0.0178, secondary eigenvalue =0.0022, the variance contribution rate of the first principal component = 0.0178 ÷ (0.0178 + 0.0022) = 0.89. ≠ The direction of the feature vector is driven by the data. ,in, For feature vectors Components in the dimension of net conductivity deviation, eigenvector The component on the net pH deviation dimension. Solving for... =−0.975. According to the normalization condition... And apply a constraint on the release response of residual ions from acidic disinfectants (i.e. >0 and <0, resulting in the normalized first principal component direction vector as [0.698, −0.716] (the conductivity component 0.698 is greater than zero, and the pH component −0.716 is less than zero, satisfying the constraint). However, for P2, the decomposition result may return either [0.698, −0.716] or [-0.698, 0.716], both of which satisfy the constraint. Further determination is needed.

[0091] When residual ions from acidic disinfectants enter the main channel of the dialysate, the increased ionization leads to higher conductivity and a decrease in pH. The phase plane trajectory extends from the origin to the fourth quadrant. The net conductivity deviation time series signal and the net pH deviation time series signal respond in opposite directions under the same stimulus. This is a physical characteristic that distinguishes the release of residual acidic disinfectant ions from other perturbations. In the scatter matrix, this corresponds to negative cross elements and the first principal component direction vector pointing to the fourth quadrant.

[0092] For P2 verification, =−0.0078, all sampling points within P2 are located in the fourth quadrant. >0 and <0, each product × All values ​​are negative. The mean of each product, formed by the cross elements of the scatter matrix, is −0.0078, which is less than zero. In the principal eigenvector returned by eigenvalue decomposition, the conductivity component is 0.698, which is greater than zero, and the pH component is −0.716, which is less than zero. All three conditions satisfy the acid disinfectant residual ion release response constraint, and the output first principal component direction vector [0.698, −0.716] and the first principal component variance contribution rate are 0.89.

[0093] Scatter matrices are constructed for trajectory segments P1 and P3 respectively, and eigenvalue decomposition is performed. The variance contribution rate of the first principal component of P1 is 0.71, and that of P3 is 0.65. Both candidate trajectory segments also satisfy the response constraint of acid disinfectant residual ion release. The corresponding first principal component direction vector and the variance contribution rate of the first principal component are output respectively.

[0094] The seal degradation identification module includes: The ion composition feature extraction unit is used to divide the absolute value of the pH net deviation component of the first principal component direction vector corresponding to the feature trajectory of the residual ion release phase plane of the seal by the conductivity net deviation component to obtain the ion response ratio of the current disinfection operation cycle; and to calculate the product of the variance contribution rate of the first principal component and the periodic exponential decay value calculated according to the difference of the disinfection operation cycle number to obtain the periodic feature confidence.

[0095] The ion response ratio and the confidence level of the periodic feature are arranged according to the period number to form the ion response ratio confidence level sequence. The dispersion of the ion response ratio is calculated by using the normalized value of the confidence level of the periodic feature as the coefficient for the ion response ratio confidence level sequence, and the historical response ratio dispersion is obtained.

[0096] Based on the confidence sequence and periodic characteristic confidence of the ion response ratio, the monotonic change trend of the ion response ratio with the order of disinfection operation cycles is calculated, and the drift trend of the ion response ratio is obtained.

[0097] The evolution characteristics of ion composition are derived from the historical response ratio dispersion and the ion response ratio drift trend.

[0098] For example, P2 is the characteristic trajectory of the residual ion release phase plane of the seal in this disinfection operation cycle (set as the n=8th disinfection operation cycle), the corresponding first principal component direction vector is [0.698,−0.716], the net conductivity deviation component is 0.698, the absolute value of the net pH deviation component is 0.716, and the ion response ratio is 0.716÷0.698=1.026.

[0099] When residual ions from the sealing component are released into the dialysate, the ratio of the increase in conductivity to the decrease in pH is determined by the type of residual ion. Acetate ions have different molar conductivity and acidification levels than citrate ions, resulting in different trajectory inclinations on the two-dimensional phase plane. The first principal component direction vector captures the main extension direction of the entire characteristic trajectory on the two-dimensional phase plane. The ratio of the absolute value of the net pH deviation to the net conductivity deviation along the main extension direction is the ion response ratio, which directly corresponds to the type of residual ion entering the dialysate. The ion response ratio for acetic acid-based disinfectants is typically between 0.85 and 1.05, while that for citrate-based disinfectants is typically between 1.20 and 1.45. In this embodiment, the ion response ratio of 1.026 falls within the acetic acid range, consistent with the type of disinfectant used in this disinfection operation.

[0100] The ion response ratios for each preceding disinfection cycle (cycles 1 to 7) were 1.02, 0.99, 1.01, 1.03, 1.05, 1.07, and 1.06, respectively. Together with the 1.026 of cycle 8, these constitute a sequence of ion response ratios for eight cycles. The exponential decay coefficient of the periodic exponential decay value was selected. The exponential decay coefficient is 0.15 (determined based on the degradation timescale of the micro-gap geometry of the seal). The seal degradation process is slow, requiring a balance between the high weighting of recent data and the statistical stability of historical data. γ=0.15 ensures an effective memory half-life of approximately 4.6 sterilization operation cycles for weight decay, covering sufficient historical information to suppress anomalies in single measurements while maintaining sufficient sensitivity to recent degradation trends. The exponential decay value for each period is With n=8, the exponential decay values ​​for periods 1 to 8 are 0.350, 0.407, 0.472, 0.549, 0.638, 0.741, 0.861, and 1.000, respectively. The variance contribution rates of the first principal component for each period are 0.83, 0.85, 0.86, 0.84, 0.87, 0.88, 0.87, and 0.89, respectively. Multiplying these by the corresponding exponential decay values ​​yields the periodicity confidence scores for periods 1 to 8 as follows: 0.290, 0.346, 0.406, 0.461, 0.556, 0.652, 0.749, and 0.890, respectively. Arranging the ion response ratios and periodicity confidence scores for the eight periods according to their period numbers constitutes an ion response ratio confidence sequence.

[0101] The sum of the confidence scores for the eight periodic features is 4.35. Dividing the confidence score of each periodic feature by 4.35 yields the normalized coefficient. The normalized coefficients for periods 1 to 8 are: 0.0667, 0.0795, 0.0933, 0.106, 0.1278, 0.1499, 0.1722, and 0.2046, respectively. Using the normalized coefficients as weighting factors, the weighted mean of the ion response ratio is calculated to be 1.037. The sum of the squares of the differences between the ion response ratio for each period and the weighted mean, multiplied by the normalized coefficients, yields the historical response ratio dispersion of 0.00057.

[0102] The ionic composition shift induced by aging of the seals does not strictly follow a linear law. The monotonic variation trend of the ion response ratio accompanying the disinfection operation cycle sequence is quantified using a weighted Kendall τ statistic. Kendall τ only measures the monotonicity of the sequence, without making parametric assumptions about the variation pattern, and uses the confidence level of the cycle characteristics of each cycle as the weight, making the contribution of recent cycles to trend judgment higher than that of long-term cycles. For the ion response ratio sequence and the corresponding cycle characteristic confidence level sequence of n disinfection operation cycles, all C(n,2) comparison pairs are enumerated. , )( < ), in the form of pre-periodic periodicity confidence level Weights: > When this pair is cooperative, increment the sign value by 1; < The time is the reverse pair, denoted by a sign value of -1. The weighted Kendall's τ is given by the formula:

[0103] ; The denominator is each preorder period. The weights are based on the number of comparisons (n−) it participates in. Weighted summation. Using pre-order periodicity. Rather than subsequent cycles As a weight, it makes the comparison between recent periods causative factors It has a higher weighting and a larger proportion, and is more sensitive to recent directional changes.

[0104] The ion response ratio sequences for the first eight disinfection operation cycles in this embodiment are 1.02, 0.99, 1.01, 1.03, 1.05, 1.07, 1.06, and 1.026, with corresponding cycle feature confidence sequences of 0.290, 0.346, 0.406, 0.461, 0.556, 0.652, 0.749, and 0.890. There are a total of C(8,2) = 28 comparison pairs, including 7 reverse pairs ((1,2), (1,3), (4,8), (5,8), (6,7), (6,8), and (7,8), and the remaining 21 pairs are co-pairs. Substituting these into the above formula yields an ion response ratio drift trend of 0.376. Substituting the full-precision values ​​of the contribution rates of each component of the principal component direction vector and the variance directly into the formula, and eliminating intermediate rounding errors, the ion response ratio drift trend is also found to be 0.376.

[0105] The evolution of ion composition during the 8th disinfection cycle was characterized by a historical response ratio dispersion of 0.00057 and an ion response ratio drift trend of 0.376.

[0106] The construction of the drift event classifier includes: Using the peak position of the distance sequence from the sampling point in each candidate fourth quadrant trajectory segment to the origin of the two-dimensional phase plane as the boundary, the monotonically increasing consistency of the distance sequence before the peak position and the monotonically decreasing consistency of the distance sequence after the peak position are calculated, and the product of the two is used to obtain the single-peak morphology index.

[0107] From historical operation records, candidate fourth quadrant trajectory segments extracted during continuous disinfection operation cycles prior to seal replacement events were labeled as positive samples of drift events; candidate fourth quadrant trajectory segments extracted during disinfection operation cycles during the initial service phase after seal replacement events were labeled as negative samples of drift events.

[0108] The positive and negative samples of the drift events are hierarchically sampled to construct a training set; the training set is trained by supervised learning using the variance contribution rate of the first principal component and the unimodal morphology index as classification features to obtain the drift event classifier.

[0109] For example, both the monotonically increasing and monotonically decreasing conformity are continuous values, represented by scalars within the interval [0,1] to reflect the degree of agreement between the distance sequences before and after the peak and the ideal monotonic form. For the distance difference between adjacent sampling point pairs in the segment before the peak, the proportion of sampling point pairs with a difference greater than zero to the total number of adjacent sampling point pairs in the segment before the peak is calculated, yielding the monotonically increasing conformity. For the distance difference between adjacent sampling point pairs in the segment after the peak, the proportion of sampling point pairs with a difference less than zero to the total number of adjacent sampling point pairs in the segment after the peak is calculated, yielding the monotonically decreasing conformity.

[0110] Taking trajectory segment P1 as an example, the distance sequence within P1 consists of 36 sampling points. The peak of the distance sequence appears at the 18th sampling point. The segment before the peak consists of 16 adjacent sampling point pairs formed by sampling points 1 to 17, and the segment after the peak consists of 18 adjacent sampling point pairs formed by sampling points 18 to 36. Among the first 16 adjacent sampling point pairs, 13 pairs have a distance difference greater than zero, with a monotonically increasing compliance rate = 13 ÷ 16 = 0.8125. Among the second 18 adjacent sampling point pairs, 16 pairs have a distance difference less than zero, with a monotonically decreasing compliance rate = 16 ÷ 18 = 0.889. The single-peak morphology index = 0.8125 × 0.889 ≈ 0.72.

[0111] It should be noted that the selection criterion of "distance sequence showing a monotonically increasing and then monotonically decreasing pattern" for inclusion in the single-peak candidate trajectory segment set uses binary judgment. The single-peak morphology index, after the single-peak candidate trajectory segment set is determined, further quantifies the quality of the single-peak morphology of each segment, providing continuous input features for the classifier. The brief disturbance segment generated in the servo control loop of the hemodialysis machine during the initial recovery phase, even if passing the single-peak binary screening, usually exhibits multiple local reversals in the preceding or following segments, resulting in low monotonicity and a single-peak morphology index significantly lower than that of the actual seal release event. P1, affected by the superposition of servo disturbances, has a single-peak morphology index of 0.72; P2, with sufficient trajectory extension and a clear single-peak structure, has a single-peak morphology index of 0.93.

[0112] Historical operation records were collected from multiple similar hemodialysis machines, covering a total of 42 seal replacement events. Candidate fourth-quadrant trajectory segments extracted within the 20 consecutive disinfection operation cycles prior to the seal replacement event were labeled as positive drift event samples, while candidate fourth-quadrant trajectory segments extracted within the first 10 disinfection operation cycles after the seal replacement were labeled as negative drift event samples. The reason for using the first 20 cycles, rather than all historical cycles, as the positive sample window is that there is an asymptomatic degradation period from normal service to the detection of residual ion release in the seal. Candidate fourth-quadrant trajectory segments in earlier cycles are mostly servo perturbations rather than actual release events, and labeling them as positive samples would introduce labeling noise. The initial service phase after replacement was used for 10 cycles because the rubber parts of the seal require several disinfection and immersion cycles to gradually stabilize after replacement. During this phase, candidate fourth-quadrant trajectory segments are pure background perturbations, and the physical basis for labeling them as negative samples is clear. A total of 374 positive drift event samples and 218 negative drift event samples were obtained.

[0113] Based on the number of disinfection operation cycles between each sample's disinfection operation cycle and the seal replacement event, the positive drift event samples were divided into three layers: 1-5 cycles, 6-12 cycles, and 13-20 cycles from the replacement event. The negative drift event samples were divided into two layers: 1-5 cycles and 6-10 cycles from the replacement event. Random sampling was performed within each layer until the number of samples in each layer was consistent, resulting in the training set. The purpose of stratified sampling is that the proportion of the cycle closest to the replacement event (end of degradation) in the positive drift event samples is naturally high. Without stratification, the classifier tends to only identify extreme release events in the end of degradation, thus suppressing its ability to identify early and mid-stage seal releases. The variance contribution rate of the first principal component of each sample and the unimodal morphology index constituted a two-dimensional classification feature. Logistic regression was used for supervised learning training on the training set. The linear combination of the two features was mapped by the sigmoid function to obtain the drift event classification confidence. Cross-entropy loss was used as the optimization objective, and L2 regularization was used to suppress overfitting, resulting in the drift event classifier. After training, the classifier returned a confidence score of 0.38 for the drift event returned by P1 and 0.87 for P2.

[0114] The seal degradation identification module includes: The equivalent volume calculation submodule is used to calculate the decay time constant of each preceding disinfection operation cycle in the degradation characteristic sequence of the seal by applying an exponential decay weight with the difference in the disinfection operation cycle number as the exponent, and then performing a weighted average calculation to obtain the micro-gap decay time constant of the nth disinfection operation cycle.

[0115] Using the micro-gap decay time constant as the decay parameter, the peak amplitude of the nth period is subjected to multi-preceding period cross-cycle residual correction to obtain the cross-cycle residual corrected peak amplitude.

[0116] The stable operating segment is defined as the time interval from the start of the current disinfection operation cycle sampling window to the start of the characteristic trajectory of the residual ion release phase plane of the seal, excluding all identified candidate fourth quadrant trajectory segments. The mean net conductivity deviation of the stable operating segment is used as the baseline conductivity. The ratio of the absolute value of the mean net pH deviation of the stable operating segment to the mean net conductivity deviation is used as the baseline ion response ratio. The ratio of the square of the baseline ion response ratio to the sum of the square of the baseline ion response ratio and the historical response ratio dispersion is used as the drift trend effective coefficient. The trend-corrected ion response ratio is obtained by summing the product of the ion response ratio drift trend and the drift trend effective coefficient with the baseline ion response ratio. The ion composition-corrected equivalent conductivity is obtained by multiplying the ratio of the baseline ion response ratio to the trend-corrected ion response ratio by the baseline conductivity.

[0117] Peak amplitude of cross-cycle residual correction Micro-gap decay time constant Baseline ion response ratio Baseline conductivity Compared with trend-corrected ion response ratio The equivalent volume index was calculated. The formula for calculating the equivalent volume index is as follows: .

[0118] in, This is the preset disinfectant ion diffusion volume equivalent coefficient.

[0119] Arranged according to the disinfection operation cycle number, a sequence of equivalent volume indexes for the micro-gap of the sealing component is formed.

[0120] For example, the eighth disinfection operation cycle of the hemodialysis machine in this embodiment is taken as an example.

[0121] The decay time constants for the first to seventh disinfection operation cycles are 312, 318, 315, 321, 325, 329, and 334 s, respectively. The decay time constants drift slowly with the degradation of the seal, and the more recent cycles are more representative of the current micro-gap geometry. A time constant decay coefficient of 0.1 was selected (determined based on the evolutionary inertia of the micro-gap geometry of the sealing component. The micro-gap geometry is dominated by the creep and wear accumulation of the sealing component material. The changes between adjacent disinfection operation cycles are small, and the measurement fluctuations of a single cycle may mask the true geometric expansion trend. A smaller value is chosen to give higher weight to recent cycles to capture the latest degradation state, while widening the weight gradient between each preceding cycle, effectively distinguishing the true evolution trend from random fluctuations within the available cycle sequence length). The weights of each preceding cycle were calculated using the difference in cycle number as a power. The unnormalized weights for cycles 1 to 7 were 0.497, 0.549, 0.607, 0.670, 0.741, 0.819, and 0.905, respectively, and the normalized weights were 0.104, 0.115, 0.127, 0.140, 0.155, 0.171, and 0.189, respectively. The weighted average yielded a micro-gap decay time constant of 323.7 s.

[0122] The residual conductivity of each preceding disinfection cycle is superimposed on the peak amplitude of the 8th disinfection cycle. After cross-cycle residual correction of multiple preceding cycles, the cross-cycle residual correction peak amplitude is 0.138 mS / cm.

[0123] The stable operation segment is the time period from the start of the current disinfection operation cycle's data acquisition window to the start of the characteristic trajectory of the residual ion release phase plane of the sealing component. Within this time period, the net conductivity deviation and net pH deviation time signals exhibit low-amplitude random fluctuations due to the fact that ion release has not yet begun, without any continuous unidirectional shift. The criterion is that the rolling average of the net conductivity deviation and net pH deviation does not exceed three times their respective sampling standard deviations. In this embodiment, P2 starts at 210 seconds after the disinfection operation ends. Sampling points falling into the candidate fourth quadrant trajectory segment P1 (82 to 100 seconds, a total of 36 sampling points) are excluded. The stable operation segment consists of 384 sampling points, from 0 to 82 seconds (164 sampling points) and from 100 to 210 seconds (220 sampling points).

[0124] In the stable operation phase, among the 384 sampling points, the mean net conductivity deviation was 0.011 mS / cm, i.e., the baseline conductivity was 0.011 mS / cm; the mean absolute value of the net pH deviation was 0.0102. The baseline ion response ratio was calculated as the ratio of the mean absolute value of the net pH deviation to the mean net conductivity deviation, resulting in a baseline ion response ratio of 0.927. Using the baseline ion response ratio of 0.927 as a benchmark, the sum of the square of the baseline ion response ratio (0.8593) and the historical response ratio dispersion (0.00057) was calculated to be 0.8599. Dividing the square of the baseline ion response ratio by 0.8599 yielded the drift trend effective coefficient (0.8593 ÷ 0.8599 = 0.9993). Finally, the trend-corrected ion response ratio (1.303) was obtained by summing the product of the ion response ratio drift trend (0.376) and the drift trend effective coefficient (0.9993) (0.3757) with the baseline ion response ratio (0.927).

[0125] Disinfectant ion diffusion volume equivalent coefficient This is a dimensionless normalization coefficient used to convert the product of the conductivity signal amplitude and the decay time constant into an equivalent volume index that monotonically corresponds to the degree of micro-gap degradation. Its value is determined through the following normalization calibration procedure: On at least five hemodialysis machines of the same model, during the initial service phase after the seal replacement, the net conductivity deviation time-series signal and the net pH deviation time-series signal were continuously collected for the first five disinfection operation cycles. The cross-cycle residual correction peak amplitude, adaptive decay time constant estimate, baseline ion response ratio, and baseline conductivity versus trend-corrected ion response ratio were extracted for each cycle using this method. The original values ​​of the equivalent volume index for each cycle were calculated (let...). The calculated value when =1), the average of the original values ​​of the first 5 cycles is taken as the initial service baseline original value of the hemodialysis machine; the arithmetic mean of the initial service baseline original values ​​of more than 5 hemodialysis machines is taken to obtain the initial service baseline original mean value. ;make Among them, the predetermined reference base value The selection principle is to make the seal failure threshold similar to the initial service reference value. The ratio is not less than 2, and the periodic fluctuation of the equivalent volume index sequence does not exceed 2. 10%. For this embodiment, 10% is selected. The value is 0.025 mL, corresponding to the disinfectant ion diffusion volume equivalent coefficient. 1.0× mL / s. When the type or nominal concentration of disinfectant changes, the molar conductivity and concentration of residual ions change, and the conductivity signal amplitude under the same degree of micro-gap degradation changes accordingly, requiring the above calibration procedure to be repeated.

[0126] The cross-cycle residual correction peak amplitude was 0.138 mS / cm, the micro-gap decay time constant was 323.7 s, the baseline ion response ratio was 0.927, the baseline conductivity was 0.011 mS / cm, the trend-corrected ion response ratio was 1.303, and the disinfectant ion diffusion volume equivalent coefficient was 1.0× Substituting mL / s into the equivalent volume index calculation formula, we obtain the equivalent volume index for this disinfection operation cycle as 0.0289 mL.

[0127] The equivalent volume indices for the first eight disinfection operation cycles are 0.0262mL, 0.0269mL, 0.0271mL, 0.0275mL, 0.0280mL, 0.0283mL, 0.0286mL, and 0.0289mL, respectively. Arranged according to the disinfection operation cycle number, they form the equivalent volume index sequence of the micro-gap of the sealing component.

[0128] The equivalent volume calculation submodule includes: The cumulative cross-cycle time interval from each preceding disinfection operation cycle to the nth disinfection operation cycle is obtained by using the difference between the end time of the disinfection operation in the nth disinfection operation cycle and the end time of the disinfection operation in each preceding disinfection operation cycle.

[0129] Using the cross-cycle residual correction peak amplitude obtained from each preceding disinfection operation cycle as the initial amplitude and the micro-gap decay time constant as the decay parameter, the exponential decay calculation of each preceding disinfection operation cycle according to the corresponding cumulative cross-cycle time interval is performed and then summed to obtain the contribution of the superimposed residual conductivity of multiple preceding cycles.

[0130] The residual corrected peak amplitude across cycles is obtained by subtracting the contribution of residual conductivity from multiple preceding cycles from the peak amplitude of the nth disinfection operation cycle in the seal degradation characteristic sequence.

[0131] For example, consider the scenario where a dialysis center performs disinfection operations continuously on the same day after equipment maintenance. Let's assume this is the n=5th disinfection operation cycle. Four disinfection operations were performed on the day the maintenance was completed. A rinsing and residue removal process is required between adjacent disinfection operations, compressing the actual operation interval to 60-90 minutes: the difference between the end time of the 4th and 5th disinfection operation cycles is 3600s (60min), the 3rd cycle is 7560s (126min), the 2nd cycle is 12240s (204min), and the 1st cycle is 16920s (282min).

[0132] The initial amplitude, rather than the observed peak amplitude, is obtained from the cross-cycle residual correction peak amplitudes of previous disinfection cycles because the observed peak amplitude itself includes the residual superposition from earlier cycles. If the observed peak amplitude is used recursively, the residual components already superimposed in previous cycles will be repeatedly included in the correction calculations of subsequent cycles, resulting in a cycle-by-cycle accumulation of error. By using the corrected cross-cycle residual correction peak amplitude as the initial amplitude, the residual attenuation estimates for each cycle are independent of each other.

[0133] The peak amplitudes of the cross-cycle residual corrections for each preceding cycle are as follows: =0.112mS / cm =0.118mS / cm =0.124 mS / cm =0.129 mS / cm. The micro-gap decay time constant is 2400 s. Using 2400 s as the decay parameter, the residual conductivity contributions of each preceding cycle at the end of the 5th disinfection cycle are as follows:

[0134] 4th cycle: 0.129× =0.0288 mS / cm; Third cycle: 0.124× =0.0053 mS / cm; Second cycle: 0.118× =0.00072 mS / cm; First cycle: 0.112 × ≈0; The contribution of residual conductivity from multiple preceding cycles is approximately 0.0288 + 0.0053 + 0.00072 + 0 ≈ 0.0348 mS / cm. The observed peak amplitude during the 5th disinfection cycle is 0.163 mS / cm, and the peak amplitude of the residual correction across cycles is... =0.163−0.0348=0.1282mS / cm.

[0135] The residual contribution of the fourth period is 0.0288 mS / cm, accounting for 17.7% of the observed peak amplitude, while the contribution of the third period is 3.3%. If the observed peak amplitude of 0.163 mS / cm is used directly in the equivalent volume calculation without cross-period residual correction, the equivalent volume index will be overestimated.

[0136] The maintenance cycle prediction module includes: The entire disinfection operation cycle sequence of the equivalent volume index of the micro-gap of the sealing component is traversed as candidate degradation inflection points. For each candidate degradation inflection point, the equivalent volume index sequence of the micro-gap of the sealing component is divided into a segment before the inflection point and a segment after the inflection point. Joint regression is performed on the segment before the inflection point and the segment after the inflection point. The two linear fitting lines are made to take the same ordinate value at the candidate degradation inflection point. The candidate degradation inflection point position corresponding to the minimum sum of the squares of the linear fitting residuals of the segment before the inflection point and the segment after the inflection point is taken as the degradation inflection point, thus obtaining the degradation rate before the inflection point and the degradation rate after the inflection point.

[0137] When the degradation rate after the inflection point is greater than the degradation rate before the inflection point, the degradation rate after the inflection point is taken as the current degradation rate; when the degradation rate after the inflection point is not greater than the degradation rate before the inflection point, the degradation rate before the inflection point is taken as the current degradation rate; the remaining number of available disinfection cycles is obtained by dividing the difference between the equivalent volume degradation judgment threshold of the seal and the end value of the equivalent volume index sequence of the micro gap of the seal by the current degradation rate.

[0138] For example, taking the equivalent volume index sequence of the micro-gap of the sealing component for the first 8 disinfection operation cycles as an example, the equivalent volume indices are 0.0262mL, 0.0269mL, 0.0271mL, 0.0275mL, 0.0280mL, 0.0283mL, 0.0286mL, and 0.0289mL, respectively. The positions of the 2nd to 7th disinfection operation cycles are used as candidate degradation inflection points, with at least one sampling point retained at the beginning and end to ensure that the segment before and after the inflection point can be linearly regressed.

[0139] A continuous piecewise linear model is applied to each candidate location, meaning that the straight line segment before and after the inflection point shares the same ordinate value at the candidate degradation inflection point. The reason for applying a continuity constraint instead of fitting the two segments independently is that the equivalent volume index describes the physical process of the continuous expansion of the micro-gap of the same seal. The equivalent volume before and after the inflection point should take the same value at the inflection point. Independent fitting allows the two segments to jump at the inflection point, which is inconsistent with the physical process and will lead to the residual sum of squares being artificially suppressed and the inflection point location estimation being unstable.

[0140] Taking the second disinfection cycle as an example, the segment before the inflection point covers cycles 1 to 2, and the segment after the inflection point covers cycles 2 to 8. Both segments share the same ordinate value of 0.0269 mL at the second cycle. Linear regression of the segment before the inflection point yields a slow degradation rate of 0.0007 mL / cycle, while linear regression of the segment after the inflection point yields an accelerated degradation rate of 0.00033 mL / cycle. After traversing all candidate locations, the sum of the squared residuals corresponding to the second disinfection cycle is the smallest, thus confirming it as the degradation inflection point. The degradation rate before the inflection point is 0.0007 mL / cycle, and the degradation rate after the inflection point is 0.00033 mL / cycle.

[0141] In this embodiment, the degradation rate after the inflection point (0.00033 mL / cycle) is less than the degradation rate before the inflection point (0.0007 mL / cycle), which does not meet the accelerated degradation confirmation condition of "the degradation rate after the inflection point is greater than the degradation rate before the inflection point". This indicates that the equivalent volume index sequence of the sealing component microgap shows a near-linear growth over the current 8 cycles and no identifiable accelerated degradation turning point has yet appeared. The degradation rate before the inflection point (0.0007 mL / cycle) is taken as the current degradation rate. The degradation threshold of the sealing component equivalent volume is determined to be 0.060 mL based on historical failure data of sealing components of similar hemodialysis machines. The end value of the equivalent volume index sequence of the sealing component microgap is 0.0289 mL. The remaining usable disinfection cycles = (0.060 − 0.0289) ÷ 0.0007 ≈ 44 disinfection operation cycles.

[0142] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based medical device maintenance prediction system, characterized in that, The system includes: The disinfection signal acquisition module is used to acquire conductivity time-series signals, pH time-series signals, dialysate temperature time-series signals, and bicarbonate concentrate proportioning pump real-time output time-series signals after the disinfection operation of the hemodialysis machine is completed, and to obtain the nominal concentration of bicarbonate concentrate. The disinfection signal decoupling module is used to perform temperature compensation on the conductivity time series signal based on the dialysate temperature time series signal to obtain a temperature-compensated conductivity time series signal; and to decouple the temperature-compensated conductivity time series signal and the pH time series signal based on the real-time output time series signal of the bicarbonate concentrate proportioning pump and the nominal concentration of the bicarbonate concentrate to obtain the net conductivity deviation time series signal and the net pH deviation time series signal. The sealing component degradation identification module is used to extract the sealing component degradation feature sequence from the time series signals of net conductivity deviation and net pH deviation of each disinfection operation cycle using machine learning methods; and calculates the equivalent volume index of the sealing component micro gap based on the sealing component degradation feature sequence to obtain the equivalent volume index sequence of the sealing component micro gap. The maintenance cycle prediction module is used to fit the degradation rate of the equivalent volume index sequence of the micro-gap of the seal to obtain the remaining number of disinfection cycles; when the remaining number of disinfection cycles is lower than the preset service cycle threshold, it outputs the seal maintenance prediction command.

2. The medical device maintenance prediction system based on machine learning according to claim 1, characterized in that, The disinfection signal decoupling module includes: Obtain the conductivity and pH settings specified in the dialysis prescription from the disinfection operation records; The real-time output signal of the bicarbonate concentrate mixing pump was converted with the nominal concentration of the bicarbonate concentrate according to the dilution ratio of the dialysate at each sampling point to obtain the real-time concentration time-series curve of bicarbonate at the dialysate outlet. Based on the theoretical relationship between the molar conductivity and concentration of bicarbonate solution at a reference temperature, the real-time concentration time-series curve of bicarbonate at the dialysate outlet is converted into a time-series curve representing the theoretical contribution of bicarbonate to the conductivity of the dialysate. The theoretical contribution time-series curve is then subtracted from the steady-state conductivity contribution corresponding to the nominal concentration of the bicarbonate concentrate to obtain the theoretical contribution deviation time-series curve. Finally, the conductivity setpoint is subtracted from the temperature-compensated conductivity time-series signal, and the theoretical contribution deviation time-series curve is subtracted to obtain the net conductivity deviation time-series signal. The real-time concentration time-series curve of bicarbonate at the dialysate outlet is converted into a time-series curve of the theoretical pH contribution of bicarbonate to the dialysate pH. The theoretical pH contribution deviation time-series curve is obtained by subtracting the steady-state pH contribution corresponding to the nominal concentration of bicarbonate concentrate from the pH theoretical contribution time-series curve. The pH setpoint and the pH theoretical contribution deviation time-series curve are then subtracted from the pH time-series signal to obtain the net pH deviation time-series signal.

3. The medical device maintenance prediction system based on machine learning according to claim 1, characterized in that, The seal degradation identification module includes: The degradation feature extraction submodule is used to map the time series signals of net conductivity deviation and net pH deviation of each disinfection operation cycle onto a two-dimensional phase plane with net conductivity deviation as the horizontal axis and net pH deviation as the vertical axis to obtain the net deviation phase plane trajectory. Machine learning methods were used to identify continuous trajectory segments that fall into the fourth quadrant in the net deviation phase plane trajectory, thus obtaining the characteristic trajectory of the phase plane for residual ion release from the seal. Within the time interval corresponding to the characteristic trajectory of residual ion release phase plane in the sealing component, the net conductivity deviation at the sampling time corresponding to the maximum distance between each sampling point and the origin of the two-dimensional phase plane is taken as the peak amplitude; the attenuation time constant is obtained by least squares fitting of the time series of net conductivity deviation after the sampling time corresponding to the maximum distance between the sampling point and the origin of the two-dimensional phase plane. Based on the overall distribution of the characteristic trajectory of residual ions released from the sealing component in the two-dimensional phase plane, the evolution characteristics of ion composition are extracted. The peak amplitude, decay time constant, and ion composition evolution characteristics are used to construct the seal degradation feature vector for the current disinfection operation cycle, and arranged according to the disinfection operation cycle number to form the seal degradation feature sequence.

4. The medical device maintenance prediction system based on machine learning according to claim 3, characterized in that, The degradation feature extraction submodule includes: The residual ion feature trajectory recognition unit is used to traverse the net deviation phase plane trajectory according to the sampling time sequence, extract continuous time segments with net conductivity deviation greater than zero and net pH deviation less than zero, and obtain candidate fourth quadrant trajectory segments. Principal component analysis was performed on the two-dimensional sampling point set composed of the net conductivity deviation and the net pH deviation within each candidate fourth quadrant trajectory segment to obtain the first principal component direction vector and the first principal component variance contribution rate; the distance sequence from each sampling point within each candidate fourth quadrant trajectory segment to the origin of the two-dimensional phase plane was calculated, and candidate fourth quadrant trajectory segments with a distance sequence that first monotonically increases and then monotonically decreases were extracted to obtain a set of single-peak candidate trajectory segments; For each candidate trajectory segment in the set of single-peak candidate trajectory segments, the variance contribution rate of the first principal component and the single-peak morphology index are input into the trained drift event classifier. The candidate trajectory segment with the highest confidence in the classification of drift events is identified as the feature trajectory of the residual ion release phase plane of the sealing component.

5. The medical device maintenance prediction system based on machine learning according to claim 4, characterized in that, The residual ion feature trajectory recognition unit includes: The net conductivity deviation and net pH deviation at each sampling time within the candidate fourth quadrant trajectory segment are used to form a two-dimensional sampling vector. The product of the two-dimensional sampling vector at each sampling time and its own transpose is averaged over all sampling times to obtain a two-dimensional phase plane scatter matrix with the origin of the two-dimensional phase plane as the reference. The eigenvalue decomposition of the two-dimensional phase plane scatter matrix is ​​performed to obtain the principal direction eigenvalues, secondary direction eigenvalues, and eigenvectors of the principal direction eigenvalues ​​and secondary direction eigenvalues. The eigenvectors corresponding to the principal direction eigenvalues ​​are taken as the first principal component direction vectors. The variance contribution rate of the first principal component is obtained by dividing the principal direction eigenvalues ​​by the sum of the principal direction eigenvalues ​​and secondary direction eigenvalues. Using the fact that the cross element of the net conductivity deviation and the net pH deviation of the two-dimensional phase plane scatter matrix is ​​less than zero, and the conductivity component and pH component of the first principal component direction vector are greater than zero as constraints on the release response of residual ions of acidic disinfectant, the eigenvalue decomposition results of the candidate fourth quadrant trajectory segment are verified, and the first principal component direction vector and the first principal component variance contribution rate that satisfy the release response constraints of residual ions of acidic disinfectant are output.

6. The medical device maintenance prediction system based on machine learning according to claim 5, characterized in that, The seal degradation identification module includes: The ion composition feature extraction unit is used to divide the absolute value of the pH net deviation component of the first principal component direction vector corresponding to the feature trajectory of the residual ion release phase plane of the seal by the conductivity net deviation component to obtain the ion response ratio of the current disinfection operation cycle; and to calculate the product of the variance contribution rate of the first principal component and the periodic exponential decay value calculated according to the difference of the disinfection operation cycle number to obtain the periodic feature confidence. The ion response ratio and the confidence scores of the periodic features are arranged according to the period number to form the ion response ratio confidence score sequence. The dispersion of the ion response ratio is calculated by using the normalized value of the confidence scores of the periodic features as coefficients for the ion response ratio confidence score sequence, and the historical response ratio dispersion is obtained. Based on the confidence sequence and periodic characteristic confidence of the ion response ratio, the monotonic change trend of the ion response ratio with the order of disinfection operation cycles is calculated, and the ion response ratio drift trend is obtained. The evolution characteristics of ion composition are derived from the historical response ratio dispersion and the ion response ratio drift trend.

7. The medical device maintenance prediction system based on machine learning according to claim 4, characterized in that, The construction of the drift event classifier includes: Using the peak position of the distance sequence from the sampling point in each candidate fourth quadrant trajectory segment to the origin of the two-dimensional phase plane as the boundary, calculate the monotonically increasing consistency of the distance sequence before the peak position and the monotonically decreasing consistency of the distance sequence after the peak position, and use the product of the two to obtain the single-peak morphology index. From historical operation records, candidate fourth quadrant trajectory segments extracted during continuous disinfection operation cycles before the seal replacement event were marked as positive samples of drift events; candidate fourth quadrant trajectory segments extracted during disinfection operation cycles during the initial service phase after the seal replacement were marked as negative samples of drift events. The positive and negative samples of the drift events are hierarchically sampled to construct a training set; the training set is trained by supervised learning using the variance contribution rate of the first principal component and the unimodal morphology index as classification features to obtain the drift event classifier.

8. The medical device maintenance prediction system based on machine learning according to claim 1, characterized in that, The seal degradation identification module includes: The equivalent volume calculation submodule is used to calculate the decay time constant of each preceding disinfection operation cycle in the degradation characteristic sequence of the seal by applying an exponential decay weight with the difference in the disinfection operation cycle number as the exponent, and then performing a weighted average calculation to obtain the micro-gap decay time constant of the nth disinfection operation cycle. Using the micro-gap decay time constant as the decay parameter, the peak amplitude of the nth period is subjected to multi-preceding period cross-cycle residual correction to obtain the cross-cycle residual corrected peak amplitude. The stable operating segment is defined as the time interval from the start of the current disinfection operation cycle sampling window to the start of the characteristic trajectory of the residual ion release phase plane of the seal, excluding all identified candidate fourth quadrant trajectory segments. The mean net conductivity deviation of the stable operating segment is used as the baseline conductivity. The ratio of the absolute value of the mean net pH deviation of the stable operating segment to the mean net conductivity deviation is used as the baseline ion response ratio. The ratio of the square of the baseline ion response ratio to the sum of the square of the baseline ion response ratio and the historical response ratio dispersion is used as the drift trend effective coefficient. The trend-corrected ion response ratio is obtained by summing the product of the ion response ratio drift trend and the drift trend effective coefficient with the baseline ion response ratio. The ion composition-corrected equivalent conductivity is obtained by multiplying the ratio of the baseline ion response ratio to the trend-corrected ion response ratio by the baseline conductivity. Peak amplitude of cross-cycle residual correction Micro-gap decay time constant Baseline ion response ratio Baseline conductivity Compared with trend-corrected ion response ratio The equivalent volume index was calculated. The formula for calculating the equivalent volume index is as follows: ； in, This is the preset disinfectant ion diffusion volume equivalent coefficient; Arranged according to the disinfection operation cycle number, a sequence of equivalent volume indexes for the micro-gap of the sealing component is formed.

9. A machine learning-based medical device maintenance prediction system according to claim 8, characterized in that, The equivalent volume calculation submodule includes: The cumulative cross-cycle time interval from each preceding disinfection operation cycle to the nth disinfection operation cycle is obtained by using the difference between the end time of the disinfection operation in the nth disinfection operation cycle and the end time of the disinfection operation in each preceding disinfection operation cycle. Using the cross-cycle residual correction peak amplitude obtained in each previous disinfection operation cycle as the initial amplitude and the micro-gap decay time constant as the decay parameter, the exponential decay calculation of each previous disinfection operation cycle is performed according to the corresponding cumulative cross-cycle time interval, and then summed to obtain the superimposed residual conductivity contribution of multiple previous cycles. The residual corrected peak amplitude across cycles is obtained by subtracting the contribution of residual conductivity from multiple preceding cycles from the peak amplitude of the nth disinfection operation cycle in the seal degradation characteristic sequence.

10. A machine learning-based medical device maintenance prediction system according to claim 1, characterized in that, The maintenance cycle prediction module includes: The entire disinfection operation cycle sequence of the equivalent volume index of the micro-gap of the sealing component is traversed as candidate degradation inflection points. For each candidate degradation inflection point, the equivalent volume index of the micro-gap of the sealing component is divided into a segment before the inflection point and a segment after the inflection point. Joint regression is performed on the segment before the inflection point and the segment after the inflection point. The two linear fitting lines are made to take the same ordinate value at the candidate degradation inflection point. The candidate degradation inflection point is the position corresponding to the minimum sum of the squares of the linear fitting residuals of the segment before the inflection point and the segment after the inflection point. The degradation rate before the inflection point and the degradation rate after the inflection point are obtained. When the degradation rate after the inflection point is greater than the degradation rate before the inflection point, the degradation rate after the inflection point is taken as the current degradation rate; when the degradation rate after the inflection point is not greater than the degradation rate before the inflection point, the degradation rate before the inflection point is taken as the current degradation rate; the remaining number of available disinfection cycles is obtained by dividing the difference between the equivalent volume degradation judgment threshold of the seal and the end value of the equivalent volume index sequence of the micro gap of the seal by the current degradation rate.

Images