Semen storage strategy optimization system and method based on hybrid neural network architecture

Abstract

The invention provides a seminal fluid storage strategy optimization system and method based on a hybrid neural network architecture, and the system comprises a data collection module which is used for obtaining temperature position vibration data, associating an offline sperm motility attenuation rate, calculating a spatial-temporal correlation degree, and constructing a coupling tensor; the time sequence prediction module is used for the LSTM to receive the temperature, the gradient, the correlation degree and the sensitive factors and predict the future temperature trend, and gsperm gating is used for filtering biological noise; the multivariable analysis module is used for carrying out Transform four-head attention analysis, focusing local activity coupling in a bag by V-head, and activating and punishing a supercooled region by cold-head; the strategy generation module is used for fusing the predicted temperature, the vitality weight, the supercooling penalty and the sensitive factor and softmax to output liquid nitrogen supply flow and the activity attenuation probability; and the closed-loop correction module is used for monthly counting the off-line survival rate, calculating a correction parameter omega, correcting the correlation degree gain and the basic flow, and realizing model locking without retraining.

Description

Technical Field

[0001] This invention relates to the field of data analysis, and in particular to a system and method for optimizing semen storage strategies based on a hybrid neural network architecture. Background Technology

[0002] In existing technologies, semen storage largely relies on traditional cryopreservation techniques, involving manual periodic monitoring of parameters such as liquid nitrogen levels and ambient temperature, with replenishment based on experience. Some systems introduce a single sensor to collect environmental data and employ simple threshold triggering mechanisms to regulate storage conditions in an attempt to maintain sample viability. These methods can generally meet the needs of small-scale storage scenarios and have become common technical solutions in the industry.

[0003] However, existing technologies have obvious limitations: on the one hand, the dynamic correlation between environmental parameters and sample quality is ignored, and single threshold control is difficult to cope with complex situations such as temperature gradient changes and vibration interference, which can easily lead to deviations in replenishment timing; on the other hand, the lack of in-depth mining of historical data and multi-factor coupling analysis makes it impossible to accurately predict the sample activity decay trend, often resulting in over-replenishment causing resource waste or under-replenishment causing sample failure, making it difficult to meet the needs of large-scale, high-precision semen storage. Summary of the Invention

[0004] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a semen storage strategy optimization system and method based on a hybrid neural network architecture.

[0005] To achieve the above-mentioned objectives of this invention, this invention provides a semen storage strategy optimization system based on a hybrid neural network architecture, comprising:

[0006] The data acquisition module is used to acquire temperature, potential and vibration data, correlate them with offline sperm motility decay rate, calculate spatiotemporal correlation and construct coupling tensor;

[0007] The timing prediction module is used to predict future temperature trends based on LSTM receiver temperature, gradient, correlation, and sensitivity factors. sperm Gated filtering of biological noise;

[0008] The multivariate analysis module is used for Transformer four-head attention analysis, with the V-head focusing on local vitality coupling within the bag and the cold-head activating to penalize supercooled regions;

[0009] The strategy generation module is used to integrate predicted temperature, vitality weight, supercooling penalty and sensitivity factors, and softmax outputs liquid nitrogen replenishment flow rate and vitality decay probability.

[0010] The closed-loop calibration module is used to calculate the calibration parameter Ω for monthly offline statistical survival rate, correct the correlation gain and basic flow, and achieve model locking without retraining.

[0011] In a preferred embodiment of the above technical solution, the data acquisition module includes:

[0012] Establish a spatiotemporally aligned output vector;

[0013] S k (t) = [T k (t), H k (t), V k (t), D k (t)] T ;

[0014] Where the superscript T stands for transpose, T k (t) represents the temperature value of the k-th sensor node at time t, obtained from the PT1000 three-in-one microsensor. k (t) represents the liquid nitrogen level height at the k-th sensor node at time t, in mm. This level is obtained through dynamic weighing at the gas-liquid interface, i.e., H. k (t) = Havg(t); V k (t) represents the effective vibration value of the k-th sensor node at time t, integrated by the MEMS accelerometer. , where a is acceleration, and D is acceleration. k (t) represents the sperm motility decay rate of the semen bag adjacent to the k-th sensor. Although it is offline data, it is associated with the corresponding spatial location k via an RFID tag. The most recent detection value before time t is completed through linear interpolation to form a time-aligned data with the real-time sensing data. This is the offline microscopic detection method D. k (t) = Δmot k / Δt,Δmot k This indicates two consecutive thawing tests.

[0015] In a preferred embodiment of the above technical solution, the data acquisition module includes:

[0016] The spatiotemporal correlation coefficient is defined as the multivariate cross-correlation coefficient between temperature, liquid level, vibration, and sperm motility decay rate.

[0017] S k (t) Decomposed into environmental factors and sample quality indicators according to physical properties:

[0018]

[0019] ;

[0020] Computational environment and spatiotemporal correlation degree R of samples i，j(t), where t represents time;

[0021] Spatiotemporal correlation between environmental parameters and sample quality degradation trajectory:

[0022] ;

[0023] Among them, R i，j (t) represents the spatiotemporal correlation between the i-th environmental parameter and the j-th sample quality index at time t, providing a basis for predictive compensation. Let be the spatial mean of the i-th environmental parameter at time t. The units are i=1: ℃, i=2: mm, i=3: mm / s; Let be the spatial mean of the quality index of the j-th sample at time t. The unit is % / h; n is the total number of spatial locations, n=9, which is dimensionless, i.e., a 3×3 array.

[0024] In a preferred embodiment of the above technical solution, the data acquisition module includes:

[0025] Generate lag compensation calculation, R' i，j (t,τ opt )= R i，j (t-τ opt )* Γ(τ opt ), τ opt The optimal lag time, reflecting the temperature and vitality response delay, is calculated using an online sliding search method. ;τ∈[0, 600s], the optimal lag time averages 4-6 minutes, enabling advance prediction of supply needs, Γ(τ opt () represents the gain due to lag compensation. , τ opt The maximum value of 1 is obtained when τ = 300, and it continues to increase with τ. opt When the value deviates from 300, the function value will decay exponentially; ρ D，T (τ) is the lag cross-correlation adjustment factor, ρ D，T (τ)=corr(D k (t), T k (t-τ)), T k (t−τ) is the temperature value measured at the k-th position at time t−τ;

[0026] Semen-liquid nitrogen interface coupling tensor The dimensions are 3×3×5, and vertical heat conduction information is preserved. The calculation method is as follows:

[0027] ;

[0028] ⨂ represents the spatial outer product operator, constructing a 3×3 spatial grid; ⊕ is the feature concatenation operator, which combines 4-dimensional S... k With 1D R' i，j The vectors are concatenated into a 5-dimensional vector; the subscripts s and l are the row and column indices of the array, and s, l∈{1,2,3} correspond to the positions of the 3×3 semen bags.

[0029] In a preferred embodiment of the above technical solution, the time series prediction module includes:

[0030] Make LSTM input X lstm (t) is:

[0031] X lstm (t) = [T(t), ∇T(t), R′ i，j (t,τ opt ), ξ(t)]; where T(t) is the current average temperature in °C, ∇T(t) represents the maximum space temperature gradient, in °C / cm. d km R′ represents the sensor spacing. i，j (t,τ opt The unit is dimensionless; ξ(t) is the semen thermo-shock sensitivity factor, also dimensionless; derived from R′ i，j With V rms Real-time calculations generate a three-channel time series consisting of correlation, gradient, and vibration.

[0032] Semen thermo-vibration sensitivity factor ξ(t) = α·|∇T(t)| + β·V rms (t) + γ·ΔT(t–60);

[0033] V rms (t) represents the effective value of vibration velocity within 10 seconds, in mm / s; ΔT(t–60) represents the temperature drop 1 minute ago, reflecting historical impact.

[0034] ;

[0035] in, To design the matrix, each row... , Let b be the inverse vector of the survival rate after thawing. i =1 / Survival i The offline regression coefficients were calculated as α=0.7±0.03, β=0.2±0.02, and γ=0.1±0.01, with a training set of 200 semen samples. ξ showed a linear negative correlation with the survival rate after thawing (R²=0.81), indirectly quantifying the sensitivity of semen to the current environment using measurable physical quantities.

[0036] Inserting sperm-specific gating into standard LSTM cells

[0037] g sperm (t) = σ(W g ·[ pH(t), Osm(t)] T +b g );

[0038] σ is the Sigmoid function. This represents the global average sperm motility decay rate. pH(t) represents the acidity or alkalinity of semen, measured monthly and written to RFID, with a safe range of 7.2-7.4; Osm(t) represents the osmotic pressure of semen, measured monthly and written to RFID, with a safe range of 285-295 mOsm / kg; the gate only opens when all three indicators deviate from the set range simultaneously, preventing non-biological noise from being written into long-term memory. The gating weight vector has initial values ​​of [0.5, 0.3, 0.2] and is updated via backpropagation using biological constraints. This is the gated bias term.

[0039] In a preferred embodiment of the above technical solution, the time series prediction module includes:

[0040] Generate constrained training loss function

[0041] ;

[0042] D opt To optimize the preservation decay rate, an empirical threshold of 0.1% / h is recommended; D For noise tolerance, a value of 0.05 allows for a 5% deviation; T The temperature gradient safety threshold is 2℃ / min; exceeding this value is considered an impact. △T grad / △t is the real-time temperature gradient change rate, calculated by the difference of ▽T(t); λ1, λ2, λ3 are the loss weights, λ1=0.6, λ2=0.25, λ3=0.15; This represents the prediction error;

[0043] Biological noise penalty;

[0044] For abnormal response stimulus;

[0045] Gradient updates are performed using the AdamW optimizer.

[0046] ;

[0047] ;

[0048] Only when When pH(t) and Osm(t) both deviate from the safe range, g sperm →1, allows abnormal writes to long-term memory; normally g sperm →0, suppress noise.

[0049] Output the temperature trend T for the next 5 minutes pred (t+5), assuming that after the LSTM model is trained, its output layer maps the internal state to the predicted temperature value through a linear function,

[0050] That is, T pred (t+5)=f(X′lstm (t))=T base (t+5)+ △T bio (t)+ △T LN2 (t), where f is the output function of the LSTM model.

[0051] T pred (t+5) is the temperature value at time t+5 predicted based on current and historical data; T base (t+5) represents the basic prediction of the hidden state of the LSTM, and △T bio (t) represents the biological thermal shock compensation term;

[0052] △T bio (t) = -K bio *ξ(t)* g sperm (t)*(T opt -T(t));

[0053] △T LN2 (t)=K ln *(q(t-τ ln )-q base ) *exp(-t / τ decay ); K bio T is the biological heat capacity coefficient. opt For optimal storage temperature, K ln τ is the liquid nitrogen impact factor. ln τ is the liquid nitrogen transport delay time. decay This is the thermal relaxation time.

[0054] In a preferred embodiment of the above technical solution, the multivariate analysis module includes:

[0055] Using a Transformer network structure, the output layer is configured to use the Softmax function to output the probability of sperm sample viability decay within a specific future time period. Furthermore, when this probability value reaches or exceeds a preset threshold, a liquid nitrogen replenishment command is automatically generated and triggered.

[0056] Standard global temperature, liquid level, and vibration environment are used for the first and second heads;

[0057] The Q, K, and V pins of the third active head (V-head) are connected to D. k (t), mask matrix M h Shielding areas with a space greater than 8cm ensures localized coupling within the same semen bag;

[0058] For spatial mask matrix, dkm is the spatial distance between sensor k and m, ensuring that attention is focused only on the inside of the same semen bag, eliminating false associations across bags;

[0059] The fourth cryogenically damaged head: The supercooled activation step function Θ (T < –130 ℃) outputs α. cold Used to penalize areas that are excessively cooled, and directly involved in the final strategy weights;

[0060] V-head output vitality weight α vitality Assuming V-head is one of the heads in a four-head attention mechanism, after calculating the attention weights, a dynamic attention head is set, resulting in multi-head output:

[0061] ;

[0062] Where AttentionV−head is the attention calculation function of V−head;

[0063] cold−head outputs an overcooling penalty α cold ;

[0064] α cold = Θ(T<–130 ℃)·sigmoid(W c *Attention cold-head +b c ); Θ is a step function, ensuring that the low-temperature region receives additional penalty weights.

[0065] In a preferred embodiment of the above technical solution, the strategy generation module includes:

[0066] Based on the output of the cold-head and the semen vibration sensitivity factor ξ(t), the formula for the probability of activity decay is defined as follows:

[0067]

[0068] in, These are the weights and biases of the probability output layer, trained using cross-entropy loss, with the label being the inverse of the offline survival rate; the upper threshold P... decay>0.6, triggering liquid nitrogen replenishment, the experiment was calibrated to a false positive rate of <5%;

[0069] The global average rate of sperm motility decline;

[0070] As αcold increases, combined with ξ(t) and The probability value σ is mapped to [0, 1] by the Sigmoid function;

[0071] Dual-time-scale stepped replenishment method, q(t) = q base ·[1 + K p ·ξ(t)]·S(τ); where, q base Based on a liquid nitrogen flow rate of 0.5 L / h -1 Matching the container's heat capacity; K p =0.4 is the sensitivity factor gain, which is a dimensionless gain calibrated through 200 sets of impact experiments. τ is the periodic phase, τ=t mod 60, in seconds, and needs to be calibrated experimentally. S(τ): dual-scale step function; if ξ(t)>0.8 → S=3, i.e., high-speed segment, lasting 10 s; 0.4<ξ(t)≤0.8 → S=1, i.e., normal speed segment, 40 s; ξ(t)≤0.4 → S=0.3, i.e., low-speed maintenance segment, 10 s.

[0072] Using a three-stage pulse with a fixed cycle of 60 seconds (high speed-normal speed-low speed) to replace the traditional PID continuous regulation, the solenoid valve only needs to be switched on and off twice, reducing hardware load and liquid nitrogen consumption by 30%, and temperature fluctuation is less than ±1℃.

[0073] In a preferred embodiment of the above technical solution, the closed-loop correction module includes:

[0074] The offline survival rate of the sample is calculated monthly and set as Survival. month Assuming this is obtained through actual storage experiments, the number N of thawed and surviving samples within one month is counted. alive With the total number of samples N total ,

[0075] Then Survival month = (N alive / N total )*100%,

[0076] Among them, Survival month This is the offline survival rate of the samples obtained monthly.

[0077] Ω=f(Survival month Survival expected ) is the function for calculating the correction parameters, where Survivalexpected This is the expected sample survival rate.

[0078]

[0079] Where Ω is the correction function calculated based on the actual monthly survival rate and the expected survival rate, and ΔS is the mean deviation of the survival rate, ΔS = Survival month - Survival expected ,

[0080] Let η be the batch survival rate variance, and η be the correction strength coefficient, whose correction magnitude is... Dilution, κ is the bias sensitivity; λ is the unexpected decay coefficient, which acts as an exponent to amplify. Impact on conservative optimization; μ is the low survival penalty coefficient, directly multiplied by The larger the variance, the stronger the penalty;

[0081] When ΔS>0: By leveraging the combined effect of the denominator and the exponent λ, the flow rate is reduced more conservatively when the variance is large; when ΔS < -0.1: an additional (1 − μ) is added. Multiplication ensures that when variance is too large, the flow rate is conservatively increased to avoid aggressive operations that exacerbate intra-batch differences. During normal fluctuations, denominator normalization maintains the stability of linear adjustment.

[0082] Monthly offline calibration Ω corrects R through survival rate feedback. i，j The overall gain of (t):

[0083] ;

[0084] To ensure that the correlation coefficient drifts synchronously with biological results, Ω is calculated offline at 00:00 on the 1st of each month and used as a constant in all real-time calculations for that month. This eliminates the need to retrain the LSTM and meets the FDA / GMP requirements for model lock-in traceability.

[0085] Ω is used for overall adjustment of q next month. base → q base new = q base ·Ω.

[0086] This invention also discloses a method for optimizing semen storage strategies based on a hybrid neural network architecture, comprising:

[0087] S1: Temperature, potential and vibration data are acquired through sensors, correlated with offline sperm motility decay rate, spatiotemporal correlation is calculated and coupling tensor is constructed;

[0088] S2, LSTM receives temperature, gradient, correlation, and sensitivity factors to predict the temperature trend over the next 5 minutes, g spermGated filtering of biological noise;

[0089] S3, Transformer four-head attention analysis, V-head focuses on local vitality coupling within the bag, cold-head activates and punishes supercooled areas;

[0090] S4 integrates predicted temperature, vitality weight, supercooling penalty and sensitivity factors, and outputs liquid nitrogen replenishment flow rate and vitality decay probability via softmax.

[0091] S5 calculates and corrects the offline survival rate of monthly statistics using parameter Ω, adjusting the correlation gain and basic traffic to achieve model locking without retraining.

[0092] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0093] Multi-dimensional data, including temperature, liquid level, vibration, and sperm motility decay rate, were collected, and a spatiotemporally aligned output vector was established. This comprehensively and accurately acquired various information about the semen storage environment. By generating hysteresis compensation calculations, the system can anticipate the potential impact of temperature shocks on motility, improving its sensitivity and predictive ability to changes in the semen storage environment. The constructed semen-liquid nitrogen interface coupling tensor ℑ(t) retains vertical heat conduction information. By concatenating the original physical field data with spatiotemporal correlation in the channel dimension, subsequent Transformers can directly call spatiotemporal correlation slices without redundant calculations, thus improving data processing efficiency.

[0094] The LSTM input integrates information from multiple factors, including the current average temperature, maximum spatial temperature gradient, spatiotemporal correlation after hysteresis compensation, and semen thermo-vibration sensitive factors, forming a three-channel time series of correlation, gradient, and vibration. This provides a more comprehensive reflection of changes in the semen storage environment and offers richer information for temperature trend prediction. A semen-specific gating mechanism, gsperm, is inserted into the standard LSTM cells. Through learning, the input data is filtered, and the gating is activated only when the global average sperm motility decay rate, pH, and osmotic pressure simultaneously deviate from a set range. This prevents non-biological noise from being written into long-term memory, improving the model's robustness and prediction accuracy.

[0095] A four-head attention mechanism is adopted, in which the third head, the viability head (V-head), ensures that attention is focused on the same sperm bag through spatial masking, avoiding false associations across bags, and outputs the viability weight αvitality; the fourth head, the cold-damage head, is activated only when the temperature is below the glass transition critical value of -130℃, and outputs the supercooling penalty αcold, which can more accurately analyze the impact of different factors on sperm samples.

[0096] A dual-time-scale stepped replenishment method is adopted. Based on the different value ranges of the semen vibration sensitivity factor ξ(t), liquid nitrogen is replenished by a three-stage pulse replenishment of high speed, normal speed and low speed within a fixed period of 60s, replacing the traditional PID continuous regulation, which reduces hardware load and liquid nitrogen consumption.

[0097] The offline survival rate of the monthly statistical samples is calculated, and the correction parameter is calculated using the correction parameter calculation function Ω to correct R. i，j (t) overall gain and base liquid nitrogen flow rate q base This ensures that the correlation and biological results drift synchronously, and there is no need for online incremental training of LSTM. Closed-loop correction can be completed with only one offline experiment per month, which meets the traceability requirements of FDA / GMP for model locking, while reducing the maintenance cost and complexity of the system.

[0098] This invention is applicable to the intelligent management of frozen samples in human sperm banks and assisted reproductive centers, with a particular focus on the dynamic optimization of environmental parameters and the prediction of sample viability decay in liquid nitrogen storage equipment.

[0099] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0100] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0101] Figure 1 This is a schematic diagram of the process of this invention. Detailed Implementation

[0102] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0103] This invention discloses a semen storage strategy optimization system and method based on a hybrid neural network architecture, specifically including:

[0104] Liquid nitrogen is added using a dynamic weighing method at the liquid nitrogen gas-liquid interface.

[0105] The outer contour is a top view of the inner liner of the storage container, with a diameter of 500 mm;

[0106] The 3×3 semen bag array is located at the geometric center, with a bag spacing of 10 mm, a bag wall thickness of 1 mm, and is made of PTFE.

[0107] Sensor nodes T / H / V1–4 output a combined signal of temperature, liquid level, and vibration; T5–8 is only used for temperature gradient supplementation sampling, with a spatial resolution ≤50mm.

[0108] The system adopts an LSTM-Transformer hybrid architecture to achieve closed-loop optimization of the bio-physical coupling.

[0109] Example 1

[0110] Reference Figure 1 This is the first embodiment of the present invention, which provides a semen storage strategy optimization system based on a hybrid neural network architecture, comprising:

[0111] The data acquisition module calculates the spatiotemporal correlation R between the environment and the samples using environment vectors. i，j (t), immediately generate R' i，j (t,τ opt At the same time, construct ℑ(t) slices for Transformer to call;

[0112] Timing prediction module, LSTM receiver [X] lstm (t)],g sperm Gated filter for biological noise, outputs temperature trend T for the next 5 minutes pred (t+5);

[0113] The multivariate analysis module uses a four-head attention mechanism with ℑ(t) as the Q, K, and V sources, and the V-head and cold-head output vitality weights α respectively. vitality With overcooling penalty α cold ;

[0114] The strategy generation module integrates T pred α vitality α cold And ξ(t), output by softmax the liquid nitrogen replenishment flow rate q(t) and the activity decay probability P decay ;

[0115] Closed-loop correction module: Monthly offline survival rate → Calculate Ω → Write back to correct R i，j (t) gain, enabling the correlation degree and biological outcome to drift synchronously on a monthly basis without retraining the LSTM.

[0116] In a preferred embodiment of the above technical solution, the data acquisition module includes:

[0117] Nine three-in-one microsensors, numbered k=1~9, were arranged on the same radial section of the inner wall of the semen storage container to collect data on temperature, liquid level, vibration, and sperm motility decay rate, and to establish a spatiotemporally aligned output vector.

[0118] Sk (t) = [T k (t), H k (t), V k (t), D k (t)] T ;

[0119] Where the superscript T stands for transpose, T k (t) represents the temperature value of the k-th sensor node at time t, obtained from the PT1000 three-in-one microsensor. k (t) represents the liquid nitrogen level height at the k-th sensor node at time t, in mm. This level is obtained through dynamic weighing at the gas-liquid interface, i.e., H. k (t) = Havg(t); V k (t) represents the effective vibration value of the k-th sensor node at time t, integrated by the MEMS accelerometer. , where a is acceleration, and D is acceleration. k (t) represents the sperm motility decay rate of the semen bag adjacent to the k-th sensor. Although it is offline data, it is associated with the corresponding spatial location k via an RFID tag. The most recent detection value before time t is completed through linear interpolation to form a time-aligned data with the real-time sensing data. This is the offline microscopic detection method D. k (t) = Δmot k / Δt,Δmot k This indicates the percentage change in viability between two consecutive thawing tests with an interval of Δt = 24 hours, where Δt is the sampling interval. This channel directly obtains the rate of biological quality degradation, avoiding indirect inference based solely on temperature.

[0120] The spatiotemporal correlation coefficient is defined as the multivariate cross-correlation coefficient between temperature, liquid level, vibration, and sperm motility decay rate.

[0121] S k (t) Decomposed into environmental factors and sample quality indicators according to physical properties:

[0122]

[0123] ;

[0124] Computational environment and spatiotemporal correlation degree R of samples i，j (t), where t represents time;

[0125] Spatiotemporal correlation between environmental parameters and sample quality degradation trajectory:

[0126] ;

[0127] Among them, R i，j(t) represents the spatiotemporal correlation between the i-th environmental parameter and the j-th sample quality index at time t, providing a basis for predictive compensation; Let be the spatial mean of the i-th environmental parameter at time t. The units are i=1: ℃, i=2: mm, i=3: mm / s; Let be the spatial mean of the quality index of the j-th sample at time t. The unit is % / h; n is the total number of spatial locations, n=9, i.e., a 3×3 array.

[0128] Generate lag compensation calculation, R' i，j (t,τ opt )= R i，j (t-τ opt )* Γ(τ opt ), τ opt The optimal lag time, reflecting the temperature and vitality response delay, is calculated using an online sliding search method. ;τ∈[0, 600s], the optimal lag time averages 4-6 minutes, enabling advance prediction of supply needs, Γ(τ opt () represents the gain due to lag compensation. , τ opt The maximum value of 1 is obtained when τ = 300, and it continues to increase with τ. opt When the value deviates from 300, the function value will decay exponentially; ρ D，T (τ) is the lag cross-correlation adjustment factor, ρ D，T (τ)=corr(D k (t), T k (t-τ)), T k (t−τ) is the temperature value measured at the k-th position at time t−τ;

[0129] By shifting the baseline time for correlation calculation forward by τ opt And multiply by the gain Γ(τ) opt This enables physical compensation for biological response delays, allowing the system to anticipate the potential impact of temperature shocks on vitality.

[0130] Semen-liquid nitrogen interface coupling tensor The dimensions are 3×3×5, and vertical heat conduction information is preserved. The calculation method is as follows:

[0131] ;

[0132] ⨂ represents the spatial outer product operator, constructing a 3×3 spatial grid; ⊕ is the feature concatenation operator, which combines 4-dimensional S... k With 1D R' i，jThe vectors are concatenated into a 5-dimensional vector; the subscripts s and l are the array row and column indices, and s, l∈{1,2,3} correspond to the 3×3 semen bag positions.

[0133] By concatenating the original physical field data with the spatiotemporal correlation in the channel dimension, the subsequent Transformer can directly call the spatiotemporal correlation slice without repeated calculation.

[0134] In a preferred embodiment of the above technical solution, the time series prediction module includes:

[0135] Make LSTM input X lstm (t) is:

[0136] X lstm (t) = [T(t), ∇T(t), R′ i，j (t,τ opt ), ξ(t)]; where T(t) is the current average temperature in °C, ∇T(t) represents the maximum space temperature gradient, in °C / cm. d km R′ represents the sensor spacing. i，j (t,τ opt The unit is dimensionless; ξ(t) is the semen thermo-shock sensitivity factor, also dimensionless; derived from R′ i，j With V rms Real-time calculations generate a three-channel time series consisting of correlation, gradient, and vibration.

[0137] Semen thermo-vibration sensitivity factor ξ(t) = α·|∇T(t)| + β·V rms (t) + γ·ΔT(t–60);

[0138] V rms (t) represents the effective value of vibration velocity within 10 seconds, in mm / s; ΔT(t–60) represents the temperature drop 1 minute ago, reflecting historical impact.

[0139] ;

[0140] in, To design the matrix, each row... , Let b be the inverse vector of the survival rate after thawing. i =1 / Survival i The offline regression coefficients were calculated as α=0.7±0.03, β=0.2±0.02, and γ=0.1±0.01, with a training set of 200 semen samples. ξ showed a linear negative correlation with the survival rate after thawing (R²=0.81), indirectly quantifying the sensitivity of semen to the current environment using measurable physical quantities.

[0141] Inserting sperm-specific gating into standard LSTM cells

[0142] g sperm (t) = σ(W g ·[ pH(t), Osm(t)] T +b g );

[0143] σ is the Sigmoid function. This represents the global average sperm motility decay rate. pH(t) represents the acidity or alkalinity of semen, measured monthly and written to RFID, with a safe range of 7.2-7.4; Osm(t) represents the osmotic pressure of semen, measured monthly and written to RFID, with a safe range of 285-295 mOsm / kg; the gate only opens when all three indicators deviate from the set range simultaneously, preventing non-biological noise from being written into long-term memory. The gating weight vector has initial values ​​of [0.5, 0.3, 0.2] and is updated via backpropagation using biological constraints. This is a gated bias term;

[0144] g sperm It is a gating mechanism in the LSTM model, which is learned and its output is used to filter the input data and remove noise.

[0145] Table 1 shows the comparison of setpoints.

[0146] Parameter name Parameter meaning unit Safety range / set value Remark pH(t) Semen pH - 7.2 - 7.4 Monthly random inspections, RFID writing Osm(t) Semen osmotic pressure mOsm / kg 285 - 295 Monthly random inspections, RFID writing α Calculation coefficients of semen thermo-vibration sensitivity factor ξ(t) (offline regression) - 0.7± 0.03 The training set consists of 200 semen samples. β Calculation coefficients of semen thermo-vibration sensitivity factor ξ(t) (offline regression) - 0.2± 0.02 The training set consists of 200 semen samples. γ Calculation coefficients of semen thermo-vibration sensitivity factor ξ(t) (offline regression) - 0.1± 0.01 The training set consists of 200 semen samples. Relationship between ξ(t) and survival rate after thawing Correlation between semen thermo-shock sensitive factors and post-thaw survival rate - Linear negative correlation, R² = 0.81 Indirectly quantifying the sensitivity of semen to the current environment using measurable physical quantities. <![CDATA[g sperm Initial value of gating weight vector]]> Initial setting of semen-specific gating weight vector - [0.5，0.3，0.2] Corresponding to [D^-(t), pH(t), Osm(t)], updates are performed via backpropagation using biological constraints.

[0147] Generate constrained training loss function

[0148]

[0149] D opt To optimize the preservation decay rate, an empirical threshold of 0.1% / h is recommended; D For noise tolerance, a value of 0.05 allows for a 5% deviation; T The temperature gradient safety threshold is 2℃ / min; exceeding this value is considered an impact. △T grad / △t is the real-time temperature gradient change rate, calculated by the difference of ▽T(t); λ1, λ2, λ3 are the loss weights, λ1=0.6, λ2=0.25, λ3=0.15; This represents the prediction error;

[0150] Biological noise penalty;

[0151] For abnormal response stimulus;

[0152] Gradient updates are performed using the AdamW optimizer.

[0153] ;

[0154] ;

[0155] Only when When pH(t) and Osm(t) both deviate from the safe range, g sperm →1, allows abnormal writes to long-term memory; normally g sperm →0, suppress noise.

[0156] Output the temperature trend T for the next 5 minutes pred (t+5), assuming that after the LSTM model is trained, its output layer maps the internal state to the predicted temperature value through a linear function,

[0157] That is, T pred (t+5)=f(X′lstm (t))=T base (t+5)+ △T bio (t)+ △T LN2 (t), where f is the output function of the LSTM model.

[0158] T pred (t+5) is the temperature value at time t+5 predicted based on current and historical data; T base (t+5) represents the basic prediction of the hidden state of the LSTM, and △T bio (t) represents the biological thermal shock compensation term;

[0159] △T bio (t) = -K bio *ξ(t)* g sperm (t)*(T opt -T(t));

[0160] △T LN2 (t)=K ln *(q(t-τ ln )-q base ) *exp(-t / τ decay ); K bio The biological heat capacity coefficient, calibrated at 0.15℃ / unit ξ; T opt The optimal storage temperature is -196℃, and Kln is the liquid nitrogen shock coefficient, which is -0.8℃·h / L; τ ln τ is the liquid nitrogen transport delay time, i.e., 30 seconds. decayThe thermal relaxation time is 120s; the output function explicitly separates data-driven prediction from physical model compensation, improving extrapolation capabilities under small sample conditions.

[0161] In a preferred embodiment of the above technical solution, the multivariate analysis module includes:

[0162] Using the Transformer network structure, the output layer is configured to use the Softmax function to output the probability of sperm sample viability decay within a specific future time period. When this probability value reaches or exceeds a preset threshold, a liquid nitrogen replenishment command is automatically generated and triggered.

[0163] Standard global temperature, liquid level, and vibration environment are used for the first and second heads;

[0164] The Q, K, and V pins of the third active head (V-head) are connected to D. k (t), mask matrix M h Shielding areas with a space greater than 8cm ensures localized coupling within the same semen bag;

[0165] For spatial mask matrix, dkm is the spatial distance between sensor k and m, ensuring that attention is focused only on the inside of the same semen bag, eliminating false associations across bags;

[0166] The fourth cryogenically damaged head: The supercooled activation step function Θ (T < –130 ℃) outputs α. cold Used to penalize areas that are excessively cooled, and directly involved in the final strategy weights;

[0167] V-head output vitality weight α vitality Assuming V-head is one of the heads in a four-head attention mechanism, after calculating the attention weights, a dynamic attention head is set, resulting in multi-head output:

[0168] ;

[0169] Where AttentionV−head is the attention calculation function of V−head;

[0170] cold−head outputs an overcooling penalty α cold ;

[0171] α cold = Θ(T<–130 ℃)·sigmoid(W c *Attention cold-head +b c ); Θ is a step function, ensuring that the low-temperature region receives additional penalty weights.

[0172] V-head uses spatial masking to ensure that attention is focused on the same semen bag, with the distance between sensors inside the bag being less than 8 cm, thus avoiding false associations across bags; cold-head is activated only when the temperature is below the glass transition critical value of -130°C to prevent overcooling.

[0173] In a preferred embodiment of the above technical solution, the strategy generation module includes:

[0174] Based on the output of the cold-head and the semen vibration sensitivity factor ξ(t), the formula for the probability of activity decay is defined as follows:

[0175]

[0176] in, These are the weights and biases of the probability output layer, trained using cross-entropy loss, with the label being the inverse of the offline survival rate; the upper threshold P... decay >0.6, triggering liquid nitrogen replenishment, the experiment was calibrated to a false positive rate of <5%;

[0177] The global average sperm motility decay rate

[0178] α cold Increase, combined with ξ(t) and The probability value σ is mapped to [0, 1] by the Sigmoid function;

[0179] Dual-time-scale stepped replenishment method, q(t) = q base ·[1 + K p ·ξ(t)]·S(τ); where, q base Based on a liquid nitrogen flow rate of 0.5 L / h -1 Matching the container's heat capacity; K p =0.4 is the sensitivity factor gain, which is a dimensionless gain calibrated through 200 sets of impact experiments. τ is the periodic phase, τ=t mod 60, in seconds, and needs to be calibrated experimentally. S(τ): dual-scale step function; if ξ(t)>0.8 → S=3, i.e., high-speed segment, lasting 10 s; 0.4<ξ(t)≤0.8 → S=1, i.e., normal speed segment, 40 s; ξ(t)≤0.4 → S=0.3, i.e., low-speed maintenance segment, 10 s.

[0180] Using a three-stage pulse with a fixed cycle of 60 seconds (high speed-normal speed-low speed) to replace the traditional PID continuous regulation, the solenoid valve only needs to be switched on and off twice, the hardware load and liquid nitrogen consumption are reduced by 30%, and the temperature fluctuation is <±1℃.

[0181] In a preferred embodiment of the above technical solution, the closed-loop correction module includes:

[0182] The offline survival rate of the sample is calculated monthly and set as Survival. month Assuming this is obtained through actual storage experiments, the number N of thawed and surviving samples within one month is counted. alive With the total number of samples N total ,

[0183] Then Survival month = (N alive / N total )*100%.

[0184] Among them, Survival month It is the offline survival rate of the samples obtained from monthly statistics.

[0185] Ω=f(Survival month Survival expected ) is the function for calculating the correction parameters, where Survival expected This is the expected sample survival rate.

[0186]

[0187] Where Ω is the correction function calculated based on the actual monthly survival rate and the expected survival rate, and ΔS is the mean deviation of the survival rate, ΔS = Survival month - Survival expected ,

[0188] Let η be the batch survival rate variance, and η be the correction strength coefficient, whose correction magnitude is... Dilution, κ is the bias sensitivity; λ is the unexpected decay coefficient, which acts as an exponent to amplify. Impact on conservative optimization; μ is the low survival penalty coefficient, directly multiplied by The larger the variance, the stronger the penalty;

[0189] When ΔS>0: By leveraging the combined effect of the denominator and the exponent λ, the flow rate is reduced more conservatively when the variance is large; when ΔS < -0.1: an additional (1 − μ) is added. Multiplication ensures that when variance is too large, the flow rate is conservatively increased, avoiding aggressive operations that exacerbate intra-batch differences. During normal fluctuations, denominator normalization maintains the stability of linear adjustment.

[0190] Monthly offline calibration Ω corrects R through survival rate feedback. i，j The overall gain of (t):

[0191] ;

[0192] To ensure that the correlation degree drifts synchronously with the biological results, Ω is calculated offline at 00:00 on the 1st of each month and used as a constant in all real-time calculations for that month. No LSTM retraining is required, which meets the traceability requirements of FDA / GMP for model lock-in.

[0193] Ω is used for overall adjustment of q next month. base → q base new = q base ·Ω;

[0194] No online incremental training of LSTM is required; closed-loop correction can be completed with only one offline experiment per month, meeting the requirements of traceability but not mandatory real-time updates.

[0195] Real-time computation of the third-order tensor ℑ(t): 3×3×5 dimensional floating-point data stream, 0.2Hz refresh rate, edge MCU RAM > 128kB, requires FPGA or ARM+M4 dual-core; hysteresis cross-correlation τ opt Online search: 0-600 s sliding window, recalculated every 6 hours, floating-point operation ≈ 0.8 GFLOPS, edge power consumption +1.2 W.

[0196] The overall beneficial effects of this invention are as follows:

[0197] Nine three-in-one microsensors were arranged on the same radial cross-section of the inner wall of the semen storage container to collect multi-dimensional data such as temperature, liquid level, vibration, and sperm motility decay rate. A spatiotemporally aligned output vector was established to comprehensively and accurately acquire various information about the semen storage environment. By generating hysteresis compensation calculations, the system can detect the potential impact of temperature shocks on motility in advance, improving its sensitivity and predictive ability to changes in the semen storage environment. The constructed semen-liquid nitrogen interface coupling tensor ℑ(t) retains the vertical heat conduction information. The original physical field data and spatiotemporal correlation are spliced ​​together in the channel dimension, allowing the subsequent Transformer to directly call the spatiotemporal correlation slice without repeated calculations, thus improving data processing efficiency.

[0198] The LSTM input integrates information from multiple factors, including the current average temperature, maximum spatial temperature gradient, spatiotemporal correlation after hysteresis compensation, and semen thermo-vibration sensitive factors, forming a three-channel time series consisting of correlation, gradient, and vibration. This provides a more comprehensive reflection of changes in the semen storage environment and offers richer information for temperature trend prediction. Semen-specific gating is inserted into standard LSTM cells. sperm By learning to filter input data, the gating is activated only when the global average sperm motility decay rate, pH, and osmotic pressure all deviate from the set range simultaneously, thus preventing non-biological noise from being written into long-term memory and improving the model's anti-interference ability and prediction accuracy.

[0199] A four-head attention mechanism is employed. The third head, the vitality head (V-head), uses spatial masking to ensure attention is focused on the same sperm pouch, avoiding spurious cross-pouch associations, and outputs a vitality weight α. The fourth head, the cold-damage head (cold-head), is activated only when the temperature is below the glass transition critical value of -130°C, and outputs a supercooling penalty α. cold This allows for a more precise analysis of the impact of different factors on sperm samples.

[0200] A dual-time-scale stepped replenishment method is adopted. Based on the different value ranges of the semen vibration sensitivity factor ξ(t), liquid nitrogen is replenished by a three-stage pulse replenishment of high speed, normal speed and low speed within a fixed period of 60s, which replaces the traditional PID continuous regulation, reduces the hardware load and liquid nitrogen consumption, and at the same time keeps the temperature fluctuation <±1℃.

[0201] The offline survival rate of the monthly statistical samples is calculated, and the correction parameter is calculated using the correction parameter calculation function Ω to correct R. i，j (t) overall gain and base liquid nitrogen flow rate q base This ensures that the correlation and biological results drift synchronously, and there is no need for online incremental training of LSTM. Closed-loop correction can be completed with only one offline experiment per month, which meets the traceability requirements of FDA / GMP for model locking, while reducing the maintenance cost and complexity of the system.

[0202] Example 2

[0203] The second embodiment of the present invention provides a method for optimizing semen storage strategies based on a hybrid neural network architecture, comprising the following steps:

[0204] The semen storage strategy optimization method based on a hybrid neural network architecture of the present invention includes:

[0205] This experiment aimed to verify the practical effect of the semen storage strategy optimization method based on a hybrid neural network architecture. Four stainless steel liquid nitrogen storage containers of the same capacity were selected and numbered as Control Group 1, Control Group 2, Experimental Group 1, and Experimental Group 2. Each container contained 50 bovine semen samples with an initial motility of 90%, with consistent dilution ratios and cryoprotectant formulations. The initial storage environment temperature was set at -196℃, and the initial liquid nitrogen level was 80%. Control Group 1 used a sensor-controlled method to monitor the liquid level and replenish liquid nitrogen daily at 9:00 AM and 5:00 PM. Control Group 2 used a single temperature sensor threshold control method, automatically replenishing liquid nitrogen when the temperature dropped below -195.5℃ at a fixed flow rate of 5 L / min. Experimental Group 1 adopted the complete technical solution of this invention, while Experimental Group 2 adopted a simplified scheme that removed the multivariate attention mechanism to verify the function of the core module.

[0206] The experiment was conducted as follows:

[0207] Sensor network deployment: Three temperature sensors with an accuracy of ±0.6℃ are installed on the upper, middle and lower layers of the inner wall of all containers. A liquid level sensor with an accuracy of ±1% is installed on the top. A vibration sensor with a range of 0-1Hz is installed on the bottom. The sample tray has a built-in miniature activity detection probe that automatically detects activity once every 2 hours. All sensor data are uploaded to the data processing terminal in real time via a wireless transmission module, with a sampling frequency of once every 10 minutes.

[0208] Data integration and modeling: The dynamic data integration units of experimental groups 1 and 2 preprocessed the historical data of the previous 15 days, including outlier removal using the 3σ criterion, normalization mapping to the [0, 1] interval, and calculation of the correlation between environmental parameters and sample quality using the spatiotemporal correlation formula. Taking the correlation between temperature and vigor as an example, data from six time points on day 10 were selected: 8:00, 12:00, 16:00, 20:00, 0:00, and 4:00. The temperature data (°C) for experimental group 1 were -196.1, -196.0, -195.9, -196.2, -196.0, and -195.8, and the vigor data (%) were 88.2, 88.5, 88.3, ​​88.0, 88.4, and 88.6. The calculated mean temperature was -196.0°C, the mean vigor was 88.3%, the sum of covariances was 0.32, the standard deviation of temperature was 0.13, and the standard deviation of vigor was 0.21. The final correlation coefficient R = 0.32 / (0.13 × 0.21) = 11.7, which was 0.96 after normalization.

[0209] Model analysis and strategy generation: The time-series prediction unit of experimental group 1, an improved LSTM, adaptively adjusts the number of hidden layer neurons to predict temperature changes over the next 24 hours. The attention mechanism and dynamic head allocation mechanism of the multivariate analysis unit calculate the contribution weights of temperature, liquid level, and vibration to vitality decay, which are 0.62, 0.25, and 0.13, respectively. After feature fusion, a replenishment strategy is generated (e.g., replenishment flow rate of 3.2 L / min at 10:00 for 15 minutes). Experimental group 2 only outputs the strategy through the time-series prediction unit, without a weight allocation step.

[0210] Control Execution and Feedback: After receiving the command, the strategy execution control module achieves step-by-step adjustment through the main flow valve with an accuracy of ±0.1L / min and the secondary pressure compensation valve with a response time of <0.5s. The execution feedback unit compares the actual value with the target value every 5 minutes. If the deviation exceeds ±0.3℃ or ±2% of the liquid level, a verification mechanism is activated. Control group 1 is operated by two technicians taking turns, while control group 2 executes according to a fixed threshold without feedback adjustment.

[0211] Data recording and visualization: The visualization platform generates real-time heat maps of environmental parameters with a spatial resolution of 0.5m×0.5m, a time step of 1 hour for the vitality decay curve, and low / medium / high warning levels. The average values ​​and cumulative statistical data of each group of key parameters are exported daily.

[0212] Table 2: Daily Average Values ​​of Environmental Parameters on Day 20

[0213] parameter Control group 1 (artificially regulated) Control group 2 (single threshold) Experimental Group 1 (Complete Protocol) Experimental group 2 (without attention mechanism) Average temperature (°C) -195.3 -195.7 -196.0 -195.8 Temperature fluctuation range (°C) 1.2 0.8 0.3 0.6 Average liquid level (%) 72.5 85.3 76.8 81.2 Vibration frequency (Hz) 0.35 0.32 0.27 0.29

[0214] Table 3: Sample quality indicators on day 30

[0215] parameter Control group 1 (artificially regulated) Control group 2 (single threshold) Experimental Group 1 (Complete Protocol) Experimental group 2 (without attention mechanism) Average sperm motility (%) 65.8 71.2 83.5 76.3 Vitality decay standard deviation (%) 4.2 3.1 1.5 2.3 Activity decay rate (% / day) 0.81 0.63 0.22 0.45 Sample pass rate (%) 72.0 84.0 98.0 89.0

[0216] Table 4: Results of Environment-Quality Correlation Calculation (Day 25)

[0217] Calculation Item Control group 1 (artificially regulated) Control group 2 (single threshold) Experimental Group 1 (Complete Protocol) Experimental group 2 (without attention mechanism) Temperature-Vigor Correlation 0.72 0.78 0.95 0.83 Liquid level-activity correlation 0.51 0.63 0.89 0.75 Vibration-Vibration Correlation 0.32 0.35 0.68 0.42 Multi-parameter comprehensive correlation - - 0.92 0.79

[0218] Table 5: Statistics on Liquid Nitrogen Supply over 30 Days

[0219] Supply indicators Control group 1 (artificially regulated) Control group 2 (single threshold) Experimental Group 1 (Complete Protocol) Experimental group 2 (without attention mechanism) Total supplies (L) 48.6 65.2 31.5 42.8 Average supply volume per single replenishment (L) 3.2 5.4 2.1 3.6 Supply times 15 12 15 12 Nitrogen consumption per unit of activity (L / %) 0.086 0.092 0.038 0.056

[0220] Table 6: Indicators of Control Precision (30-day average)

[0221] Accuracy indicators Control group 1 (artificially regulated) Control group 2 (single threshold) Experimental Group 1 (Complete Protocol) Experimental group 2 (without attention mechanism) Temperature control deviation (°C) 0.85 0.52 0.18 0.35 Liquid level control deviation (%) 6.3 4.8 1.2 3.5 Policy execution delay (s) 180 30 5 15 Deviation correction time (s) 300 120 20 60

[0222] Table 7: Early Warning and Response Effectiveness

[0223] Response metrics Control group 1 (artificially regulated) Control group 2 (single threshold) Experimental Group 1 (Complete Protocol) Experimental group 2 (without attention mechanism) Low-risk early warning accuracy rate (%) 68.5 72.3 96.8 85.2 Accuracy rate of medium-risk early warning (%) 75.2 69.8 97.5 88.6 High-risk warning accuracy rate (%) 82.3 78.5 99.2 91.3 False alarm rate (%) 15.6 22.4 2.1 8.7

[0224] Table 2 shows that, in terms of environmental stability, the average temperature of experimental group 1 was closest to the target value, with temperature fluctuations only 25% of those of control group 1, demonstrating the precise ability of the hybrid neural network to capture subtle temperature changes. Although control group 2 maintained a lower temperature through threshold control, excessive replenishment led to a persistently high liquid level, which in turn caused slight vibrations due to increased liquid nitrogen evaporation. In contrast, experimental group 1, through multivariate analysis, identified a latent correlation between high liquid level and high vibration. By dynamically adjusting the liquid level to 76.8%, the vibration frequency decreased to 0.27 Hz, proving that spatiotemporal correlation modeling can reveal environmental coupling relationships neglected by traditional methods.

[0225] Table 3 shows the sample quality data. The average sperm motility in experimental group 1 was 26.9% higher than that in control group 1, and the rate of motility decay was only 27.2% of that in control group 1. Furthermore, the standard deviation of motility decay was significantly lower than in other groups, indicating that this invention can maintain uniform sample quality. In control group 2, due to rigid temperature control, some samples suffered cell membrane damage from prolonged exposure to excessively low temperatures, resulting in motility failing to exceed 72%. Although experimental group 2 was superior to the control group, it lacked a multivariate attention mechanism, making it unable to distinguish the weight differences between temperature and vibration, resulting in a 7.2% lower motility compared to experimental group 1. This verifies the necessity of the attention mechanism in feature selection.

[0226] Table 4 shows the correlation calculation results, which intuitively demonstrate the technological breakthrough of this invention: the temperature-activity correlation of experimental group 1 is 31.9% higher than that of control group 1, and the multi-parameter comprehensive correlation is a quantitative indicator that traditional methods cannot achieve. This result stems from the joint modeling of timestamps and spatial locations by the dynamic data integration unit, which solves the shortcomings of isolated parameter analysis in existing technologies and provides a scientific basis for strategy generation.

[0227] Table 5 shows that in terms of resource utilization efficiency, the total replenishment of experimental group 1 was reduced by 51.7% compared to control group 2, and the nitrogen consumption per unit of vitality was only 41.3% of that of control group 2, demonstrating a significant energy-saving advantage. Due to the lag in artificial replenishment, control group 1 had a 15% redundant replenishment; the fixed threshold triggering mechanism of control group 2 resulted in "full replenishment upon reaching the threshold," with an ineffective consumption rate as high as 22.4%; experimental group 1 achieved on-demand resource allocation through short-term trend prediction by the time-series prediction unit and accurate demand calculation by the multivariate analysis unit.

[0228] Table 6 further highlights the innovation of this invention by demonstrating improved control precision and response speed: the temperature control deviation of experimental group 1 was only 21.2% of that of control group 1, and the deviation correction time was shortened by 83.3% compared to control group 2. This is attributed to the synergistic effect of the multi-stage interlocking control mechanism—the primary flow valve ensures basic precision, the secondary pressure compensation valve offsets pressure fluctuations within the container, and the emergency module achieves millisecond-level response in the event of sudden anomalies. Control group 1 often misses the optimal control opportunity due to physiological delays caused by manual operation; control group 2 lacks a feedback verification mechanism and cannot correct flow deviations caused by changes in liquid nitrogen purity.

[0229] Table 7 shows the early warning response results. The high-risk early warning accuracy and low false alarm rate of Experimental Group 1 far exceeded those of traditional methods. The core of this was the risk assessment model based on probability calculation, rather than the "either / or" threshold judgment of Control Group 2. For example, on day 28, Experimental Group 1 detected a slight increase of 0.2℃ in the temperature of a certain area within one hour. Combined with the characteristic of a synchronous increase in vibration frequency, it issued a medium-risk early warning 3 hours in advance. After intervention, a significant decrease in the sample viability in that area was avoided. In contrast, Control Group 2 did not trigger an early warning because it did not reach the temperature threshold, and the sample viability in that area ultimately decreased by 4.2%.

[0230] In summary, this invention systematically solves the problems of extensive environmental control, uneven sample quality, serious resource waste, and delayed risk warning in existing semen storage technologies by utilizing the global perception of multi-source sensor networks, the deep analysis of hybrid neural networks, and the precise execution of multi-level control. It optimizes the early warning response through risk assessment, reducing the risk of sample failure. Experimental data fully demonstrates that this method achieves a qualitative leap in storage effect, resource efficiency, and reliability, and has significant practical value.

[0231] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A semen storage strategy optimization system based on a hybrid neural network architecture, characterized in that, The system includes: The data acquisition module is used to acquire temperature, potential and vibration data, correlate them with offline sperm motility decay rate, calculate spatiotemporal correlation and construct coupling tensor; The timing prediction module is used to receive temperature, gradient, correlation, and sensitivity factors from the LSTM and predict future temperature trends. The gsperm gate filters biological noise. The multivariate analysis module is used for Transformer four-head attention analysis, with the V-head focusing on local vitality coupling within the bag and the cold-head activating to penalize supercooled regions; The strategy generation module is used to integrate predicted temperature, vitality weight, supercooling penalty and sensitivity factors, and softmax outputs liquid nitrogen replenishment flow rate and vitality decay probability. The closed-loop calibration module is used to calculate the calibration parameter Ω for monthly offline statistical survival rate, correct the correlation gain and basic flow, and achieve model locking without retraining.

2. The semen storage strategy optimization system based on a hybrid neural network architecture as described in claim 1, characterized in that: The data acquisition module includes: Establish a spatiotemporally aligned output vector; S k (t) = [T k (t),H k (t),V k (t),D k (t)] T ; Where the superscript T stands for transpose, T k (t) represents the temperature value of the k-th sensor node at time t, obtained from the PT1000 three-in-one microsensor. k (t) represents the liquid nitrogen level height at the k-th sensor node at time t, in mm. This level is obtained through dynamic weighing at the gas-liquid interface, i.e., H. k (t) = Havg(t); V k (t) represents the effective vibration value of the k-th sensor node at time t, integrated by the MEMS accelerometer. , where a is acceleration, and D is acceleration. k (t) represents the sperm motility decay rate of the semen bag adjacent to the k-th sensor. Although it is offline data, it is associated with the corresponding spatial location k via an RFID tag. The most recent detection value before time t is completed through linear interpolation to form a time-aligned data with the real-time sensing data. This is the offline microscopic detection method D. k (t) = Δmotk / Δt, where Δmotk represents two consecutive thawing detections.

3. The semen storage strategy optimization system based on a hybrid neural network architecture as described in claim 1, characterized in that: The data acquisition module includes: The spatiotemporal correlation coefficient is defined as the multivariate cross-correlation coefficient between temperature, liquid level, vibration, and sperm motility decay rate. Sk(t) is decomposed into environmental factors and sample quality indicators according to physical properties: ； ； Computational environment and spatiotemporal correlation degree R of samples i，j (t), where t represents time; Spatiotemporal correlation between environmental parameters and sample quality degradation trajectory: ； Among them, R i，j (t) represents the spatiotemporal correlation between the i-th environmental parameter and the j-th sample quality index at time t, providing a basis for predictive compensation; E i,k (t) represents the i-th type of environmental parameter value at the k-th spatial location at time t, Q j,k (t) represents the quality index value of the j-th sample at the k-th spatial location at time t. Let be the spatial mean of the i-th environmental parameter at time t. , Let be the spatial mean of the quality index of the j-th sample at time t. , where n is the total number of spatial locations.

4. The semen storage strategy optimization system based on a hybrid neural network architecture as described in claim 1, characterized in that, The data acquisition module includes: Generate lag compensation calculation, R' i，j (t, τopt) = R i，j (t-τ opt )* Γ(τ opt ), τ opt The optimal lag time, reflecting the temperature and vitality response delay, is calculated using an online sliding search method. ;τ∈[0, 600s], to achieve advance prediction of supply demand, Γ(τopt) is the lag compensation gain, ;ρ D，T (τ) is the lag cross-correlation adjustment factor, ρ D，T (τ)=corr(D k (t), T k (t-τ)), T k (t−τ) is the temperature value measured at the k-th position at time t−τ; Semen-liquid nitrogen interface coupling tensor The dimensions are 3×3×5, and vertical heat conduction information is preserved. The calculation method is as follows: ； ⨂ represents the spatial outer product operator, constructing a 3×3 spatial grid; ⊕ is the feature concatenation operator, which combines 4-dimensional S... k With 1D R' i，j The vectors are concatenated into a 5-dimensional vector; the subscripts s and l are the row and column indices of the array, and s, l∈{1,2,3} correspond to the positions of the 3×3 semen bags.

5. The semen storage strategy optimization system based on a hybrid neural network architecture as described in claim 1, characterized in that, The time series prediction module includes: Make LSTM input X lstm (t) is: X lstm (t) = [T(t), ∇T(t), R′ i，j (t,τ opt ), ξ(t)]; where T(t) is the current average temperature, ;∇T(t) represents the maximum spatial temperature gradient. d km R′ represents the sensor spacing. i，j (t,τ opt The unit is dimensionless; ξ(t) is the semen thermo-vibration sensitivity factor; derived from R′ i，j With V rms Real-time calculations generate a three-channel time series consisting of correlation, gradient, and vibration. Semen thermo-vibration sensitivity factor ξ(t) = α·|∇T(t)| + β·V rms (t) + γ·ΔT(t–60); ； in, To design the matrix, each row... , The inverse vector of the survival rate after thawing is calculated using α=0.7±0.03, β=0.2±0.02, and γ=0.1±0.01 as offline regression coefficients. The training set consists of 200 semen samples. ξ is linearly negatively correlated with the survival rate after thawing, and the sensitivity of semen to the current environment is indirectly quantified using measurable physical quantities. Inserting sperm-specific gating into standard LSTM cells g sperm (t) = σ(W g ·[ , pH(t), Osm(t)]T+b g ); σ is the Sigmoid function. This represents the global average sperm motility decay rate. pH(t) represents the acidity or alkalinity of semen; Osm(t) represents the osmotic pressure of semen. The gating weight vector is updated via backpropagation using biological constraints. This is the gated bias term.

6. The semen storage strategy optimization system based on a hybrid neural network architecture as described in claim 1, characterized in that, The time series prediction module includes: Generate constrained training loss function ； D opt To optimize the preservation decay rate, an empirical threshold of 0.1% / h is recommended; D For noise tolerance; ϵ T The temperature gradient safety threshold; ΔT grad / Δt represents the real-time temperature gradient change rate; λ1, λ2, and λ3 are the loss weights. This represents the prediction error; Biological noise penalty; For abnormal response stimulus; Gradient updates are performed using the AdamW optimizer. ； ； Only when When pH(t) and Osm(t) both deviate from the safe range, g sperm →1, allows abnormal writes to long-term memory; normally g sperm →0, suppress noise. Output the temperature trend Tpred(t+5) for the next 5 minutes. Assume that after training, the output layer of the LSTM model maps the internal state to the temperature prediction value through a linear function. That is, T pred (t+5)=f(X′ lstm (t))=T base (t+5)+ △T bio (t)+ △T LN2 (t), where f is the output function of the LSTM model. T pred (t+5) is the temperature value at time t+5 predicted based on current and historical data; T base (t+5) represents the basic prediction of the hidden state of the LSTM, and △T bio (t) represents the biological thermal shock compensation term; △T bio (t)= - K bio *ξ(t)* g sperm (t)*(T opt -T(t)); △T LN2 (t)=K ln *(q(t-τ ln )-q base ) *exp(-t / τ decay ); K bio T is the biological heat capacity coefficient. opt For optimal storage temperature, K ln τ is the liquid nitrogen impact factor. ln τ is the liquid nitrogen transport delay time. decay This is the thermal relaxation time.

7. The semen storage strategy optimization system based on a hybrid neural network architecture as described in claim 1, characterized in that, The multivariate analysis module includes: Using a Transformer network structure, the output layer is configured to use the Softmax function to output the probability of sperm sample viability decay within a specific future time period. Furthermore, when this probability value reaches or exceeds a preset threshold, a liquid nitrogen replenishment command is automatically generated and triggered. Standard global temperature, liquid level, and vibration environment are used for the first and second heads; The Q, K, and V pins of the third active head (V-head) are connected to D. k (t), mask matrix M h Shielding areas with a space greater than 8cm ensures localized coupling within the same semen bag; For spatial mask matrix, d km The spatial distance between sensors k and m ensures that attention is focused only on the inside of the same semen bag, eliminating false associations across bags; The fourth cryogenically damaged head: The supercooled activation step function Θ (T < –130 ℃) outputs α. cold Used to penalize areas that are excessively cooled, and directly involved in the final strategy weights; V-head output vitality weight α vitality Assuming V-head is one of the heads in a four-head attention mechanism, after calculating the attention weights, a dynamic attention head is set, resulting in multi-head output: ； Attention V−head It is the attention calculation function for V-head; cold−head outputs an overcooling penalty α cold ; α cold = Θ(T<–130 ℃)·sigmoid(W c *Attention cold-head +b c ); Θ is a step function, ensuring that the low-temperature region receives additional penalty weights.

8. The semen storage strategy optimization system based on a hybrid neural network architecture as described in claim 1, characterized in that, The strategy generation module includes: Based on the output of the cold-head injury and the semen vibration sensitivity factor ξ(t), the formula for the probability of activity decay is defined as follows: ； in, These are the weights and biases of the probability output layer, trained using cross-entropy loss, with the label being the inverse of the offline survival rate; the upper threshold P... decay >0.6, triggering liquid nitrogen replenishment, the experiment was calibrated to a false positive rate of <5%; The global average rate of sperm motility decline; α cold Increase, combined with ξ(t) and The probability value σ is mapped to [0, 1] by the Sigmoid function; Dual-time-scale stepped replenishment method, q(t) = q base ·[1 + K p ·ξ(t)]·S(τ); where qbase is the basic liquid nitrogen flow rate, matched with the container heat capacity; Kp is the sensitivity factor gain, which is a dimensionless gain; τ is the periodic phase; and S(τ) is the two-scale step function.

9. The semen storage strategy optimization system based on a hybrid neural network architecture as described in claim 1, characterized in that, The closed-loop correction module includes: The offline survival rate of the sample is calculated monthly and set as Survival. month Assuming this is obtained through actual storage experiments, the number N of thawed and surviving samples within one month is counted. alive With the total number of samples N total , Then Survival month = (N alive / N total )*100%, Among them, Survival month This is the offline survival rate of the samples obtained monthly. Ω=f(Survival month Survival expected ) is the function for calculating the correction parameters, where Survival expected This is the expected sample survival rate. ； Where Ω is the correction function calculated based on the actual monthly survival rate and the expected survival rate, and ΔS is the mean deviation of the survival rate, ΔS = Survival month - Survival expected , Let η be the batch survival rate variance, and η be the correction strength coefficient, whose correction magnitude is... Dilution, κ is the bias sensitivity; λ is the unexpected decay coefficient, which acts as an exponent to amplify. Impact on conservative optimization; μ is the low survival penalty coefficient, directly multiplied by The larger the variance, the stronger the penalty; When ΔS>0: By leveraging the combined effect of the denominator and the exponent λ, the flow rate is reduced more conservatively when the variance is large; when ΔS < -0.1: an additional (1 − μ) is added. Multiplication ensures that when variance is too large, the flow rate is conservatively increased to avoid aggressive operations that exacerbate intra-batch differences. During normal fluctuations, denominator normalization maintains the stability of linear adjustment. Monthly offline calibration Ω corrects R through survival rate feedback. i，j The overall gain of (t): ； Ω is used for overall adjustment of q next month. base → q base new = q base ·Ω.

10. A method for optimizing semen storage strategies based on a hybrid neural network architecture, characterized in that, include: S1: Temperature, potential and vibration data are acquired through sensors, correlated with offline sperm motility decay rate, spatiotemporal correlation is calculated and coupling tensor is constructed; S2, LSTM receives temperature, gradient, correlation and sensitivity factors to predict temperature trend in the next 5 minutes, gsperm gated filter for biological noise; S3, Transformer four-head attention analysis, V-head focuses on local vitality coupling within the bag, cold-head activates and punishes supercooled areas; S4 integrates predicted temperature, vitality weight, supercooling penalty and sensitivity factors, and outputs liquid nitrogen replenishment flow rate and vitality decay probability via softmax. S5 calculates and corrects the offline survival rate of monthly statistics using parameter Ω, adjusting the correlation gain and basic traffic to achieve model locking without retraining.

Images