A control method and system for a neural network-based separation device for noctiluca scintillans

Abstract

The present application relates to the technical field of noctiluca separation control, and particularly relates to a control method and system of a noctiluca separation device based on a neural network, comprising the following steps: based on the control separation device, a low-intensity periodic disturbance is applied to the noctiluca-containing to-be-separated liquid, so that the noctiluca generates reversible luminescence response and slow floating and sinking response in the separation channel; noctiluca response separation bands are formed according to the reversible luminescence response and the slow floating and sinking response; and a neural network control model is input based on the noctiluca response separation bands, and the neural network control model distinguishes intact noctiluca, damaged noctiluca, bubbles and non-luminescent impurities. The present application constructs noctiluca response separation bands through the reversible luminescence response and the slow floating and sinking response generated by the noctiluca after the low-intensity periodic disturbance, so that the separation objects no longer rely on single appearance recognition or fixed flow rate screening, and the pre-differentiation of intact noctiluca, damaged individuals, bubbles and non-luminescent impurities is realized based on the noctiluca living body response characteristics.

Description

Technical Field

[0001] This invention relates to the field of Noctiluca scintillans separation and control technology, and in particular to a control method and system for a Noctiluca scintillans separation device based on a neural network. Background Technology

[0002] Noctiluca scintillans is a type of planktonic organism with large cell volume, soft structure, and bioluminescent response. It is often necessary to separate it in vivo in red tide monitoring, marine ecological research, bioluminescence mechanism analysis, and algal culture. Existing separation methods mostly use sieve filtration, centrifugation sedimentation, manual aspiration, or ordinary microfluidic sorting. However, the cell membrane of Noctiluca scintillans is relatively fragile and is easily deformed, ruptured, or inactivated after being subjected to strong shear, negative pressure aspiration, or wall friction, resulting in a decrease in the integrity rate after separation.

[0003] Actual water samples often contain air bubbles, silt flocs, non-luminescent impurities, and damaged Noctiluca scintillans. Some impurities are similar to Noctiluca scintillans in size, floating trajectory, or reflective state, making it difficult to reliably distinguish them using only a fixed flow rate, a fixed separation port, or a single-frame image. Especially during continuous separation, Noctiluca scintillans exhibits dynamic changes such as reversible luminescence, slow floating and sinking, wall-adhering migration, and response decay after being disturbed. If the equipment cannot promptly pause the disturbance, adjust the suction port position, compensate the escort flow rate, and control the timing of the collection valves based on this dynamic response, it can easily lead to the target being missed, impurities being mixed in, or damage to Noctiluca scintillans.

[0004] Therefore, there is an urgent need for a separation device control method that constructs a separation zone based on the reversible luminescence response and slow floating and sinking response of Noctiluca scintillans, and utilizes neural networks to collaboratively control the disturbance, escort, aspiration and collection processes. Summary of the Invention

[0005] This invention provides a control method and system for a separation device of Noctiluca scintillans based on a neural network.

[0006] Firstly, a control method for a separation device of Noctiluca scintillans based on a neural network includes the following steps:

[0007] The separation device applies low-intensity periodic disturbances to the liquid to be separated containing Noctiluca scintillans, causing the Noctiluca scintillans to produce a reversible luminescence response and a slow floating and sinking response within the separation channel.

[0008] Based on the reversible luminescence response and slow buoyancy response, a separation zone for Noctiluca scintillans response is formed;

[0009] Based on the Noctiluca scintillans response separation zone input neural network control model, the neural network control model distinguishes between intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles, and non-luminescent impurities;

[0010] Obtain a response-holding separation control command, which is used to determine the disturbance pause time, the separation inlet alignment position, the buffer escort flow rate, and the collection valve opening sequence;

[0011] Based on the aforementioned response-maintaining separation control command, the disturbance unit, separation inlet, buffer pump, and collection valve are controlled to operate in a coordinated manner.

[0012] When the Noctiluca scintillans response separation zone remains intact and does not adhere to the wall, the intact Noctiluca scintillans is introduced into the collection branch to obtain the Noctiluca scintillans separation solution.

[0013] Secondly, a control system for a neural network-based separation device for Noctiluca scintillans includes the following modules:

[0014] The disturbance response construction module is used to control the disturbance unit to apply low-intensity periodic disturbance to the liquid to be separated containing Noctiluca scintillans, so that Noctiluca scintillans generates a reversible luminescence response and a slow floating and sinking response in the separation channel, and forms a Noctiluca scintillans response separation zone based on the reversible luminescence response and the slow floating and sinking response.

[0015] The response data acquisition module is used to acquire multiple frames of two-dimensional spatial distribution images or light intensity-displacement coupling data of the Noctiluca scintillans response separation zone within a continuous time window, and form a three-dimensional spatiotemporal feature tensor according to the acquisition sequence.

[0016] The neural network control module is used to input the three-dimensional spatiotemporal feature tensor into the neural network control model, distinguish between intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles and non-luminescent impurities, output the morphology retention coefficient of intact Noctiluca scintillans, and generate response-maintaining separation control commands.

[0017] The laminar flow escort execution module is used to control the disturbance unit to stop outputting periodic driving force according to the disturbance pause time in the response-maintaining separation control command, and to control the buffer pump to form a laminar flow escort flow field and a radial anti-wall flow field.

[0018] The inhalation port adjustment module is used to control the separation inhalation port to align with the center of gravity of the complete Noctiluca scintillans distribution according to the separation inhalation port alignment position in the response-maintaining separation control command, and to dynamically adjust the inhalation rate according to the instantaneous change rate of the morphology retention coefficient.

[0019] The collection valve control module is used to control the collection valve to open or close according to the collection valve opening sequence in the response-maintaining separation control command, so that the liquid flow containing intact Noctiluca scintillans introduced through the separation inlet enters the collection branch.

[0020] An instability protection module is used to monitor the luminescence intensity decay and buoyancy velocity amplitude of the Noctiluca scintillans response separation zone during the collection process. When the duration of the luminescence intensity decay or buoyancy velocity amplitude deviating from the preset stable range exceeds the allowable deviation time, the Noctiluca scintillans response separation zone is determined to be unstable, and the separation inlet is paused, the collection valve is closed, and the Noctiluca scintillans response separation zone is rebuilt, until the preset total collection volume or preset collection time is met before the Noctiluca scintillans separation liquid is output.

[0021] The beneficial effects of this invention are:

[0022] This invention constructs a Noctiluca scintillans response separation zone by utilizing the reversible luminescence response and slow floating and sinking response generated by Noctiluca scintillans after being subjected to low-intensity periodic disturbance. This allows the separation of objects to no longer rely on single appearance identification or fixed flow rate sieving, but to achieve pre-distinction of whole Noctiluca scintillans, damaged individuals, bubbles and non-luminescent impurities based on the living response characteristics of Noctiluca scintillans.

[0023] This invention generates response-maintaining separation control commands based on a neural network control model. It transforms the morphology retention coefficient, distribution centroid, distance to interfering objects, and arrival time at the collection port of intact Noctiluca scintillans into disturbance pause, laminar flow escort, anti-wall adhesion, inlet alignment, and collection valve timing control, thereby achieving low-damage and interference-resistant automated separation. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the present invention;

[0026] Figure 2 This is a diagram of the neural network control model architecture of Embodiment 1 of the present invention;

[0027] Figure 3 This is a schematic diagram of the system modules in Embodiment 2 of the present invention. Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0029] Example 1

[0030] like Figures 1-2 As shown, a control method for a neural network-based separation device for Noctiluca scintillans includes the following steps:

[0031] S1, the separation device applies low-intensity periodic disturbance to the liquid to be separated containing Noctiluca scintillans, so that Noctiluca scintillans generates a reversible luminescence response and a slow floating and sinking response in the separation channel, and forms a Noctiluca scintillans response separation zone based on the reversible luminescence response and the slow floating and sinking response.

[0032] S11, after the liquid containing Noctiluca scintillans is introduced into the separation channel, the disturbance unit of the separation device applies low-intensity periodic disturbance to the liquid to be separated in the separation channel; the disturbance unit can be a flexible diaphragm, a piezoelectric micro-vibrator, or a low-amplitude reciprocating liquid pushing structure. Its function is not to forcibly mix the liquid to be separated, but to make Noctiluca scintillans generate identifiable restricted floating and sinking movements in the separation channel without damaging the integrity of the Noctiluca scintillans cell membrane.

[0033] The periodic driving force output by the disturbance unit is expressed in sinusoidal form as follows:

[0034] ;

[0035] in, Indicates the time of the disturbance element. The periodic driving force output; Indicates the amplitude of the periodic driving force; The frequency of the periodic driving force is represented, and its value range is [value range missing]. ; Indicates the initial phase; The disturbance time is determined based on the stable period required for the formation of the Noctiluca scintillans response separation zone. It is usually set to 3 to 10 consecutive disturbance cycles, preferably 5 to 8 disturbance cycles. When the periodic driving force frequency is 0.5 Hz to 5 Hz, the disturbance time can be controlled between 2 s and 20 s, preferably 4 s to 12 s. If the disturbance time is too short, the reversible luminescence response and slow floating and sinking response of Noctiluca scintillans have not yet formed a stable periodic correspondence, making it difficult to accurately calibrate the Noctiluca scintillans response separation zone. If the disturbance time is too long, it is easy to cause fatigue response, wall adhesion aggregation, or local flow field accumulation and shift, affecting the subsequent separation stability.

[0036] To prevent irreversible damage to Noctiluca scintillans during disturbance, the fluid shear force corresponding to the periodic driving force must not exceed the elastic deformation threshold of the Noctiluca scintillans cell membrane, expressed as:

[0037] ;

[0038] in, Indicates at time Fluid shear force formed by the periodic driving force acting on the liquid to be separated; This represents the elastic deformation threshold of the Noctiluca scintillans cell membrane, which is used to characterize the maximum permissible shear force that allows the Noctiluca scintillans cell membrane to maintain elastic recovery without rupture or inactivation.

[0039] The elastic deformation threshold of *Noctiluca scintillans* cell membrane was obtained through a pre-calibration experiment. *Noctiluca scintillans* samples from the same batch or similar culture conditions were placed in a calibration channel identical to or scaled down from the separation equipment. The perturbation unit or microfluidic pump was controlled to gradually increase the perturbation intensity from a low shear condition. At each perturbation intensity level, the morphological recovery, luminescence recovery, and survival integrity of *Noctiluca scintillans* were recorded. When *Noctiluca scintillans* could recover approximately its original outline after perturbation, the luminescence response remained reversible, and there was no cell membrane rupture, leakage of contents, continuous collapse, or inactivation, the shear force corresponding to that perturbation level was considered to be within the elastic deformation range. When the perturbation intensity was further increased, and *Noctiluca scintillans* exhibited irreversible stretching, outline rupture, disappearance of luminescence response, or a significant decrease in integrity, the maximum shear force corresponding to the previous safe perturbation condition was determined as the elastic deformation threshold of the *Noctiluca scintillans* cell membrane.

[0040] Under low-intensity periodic disturbances, Noctiluca scintillans is affected by buoyancy, fluid resistance and periodic driving force, and it generates restricted floating and sinking motion in the vertical direction within the separation channel; the restricted floating and sinking motion is not random drift, but has a phase correlation with the periodic disturbance.

[0041] S12, during the entire process of applying low-intensity periodic perturbation, the photoelectric detection array arranged around the separation channel is activated to continuously detect the luminescence of different spatial regions in the liquid to be separated; the photoelectric detection array is arranged in sections along the length and height of the separation channel to obtain the photon count or luminescence intensity value of each region during the continuous perturbation process.

[0042] For the first in the separation channel The detection area, in the first The luminous intensity value at each sampling time is represented as:

[0043] ;

[0044] in, Indicates the first The detection area in the first The luminous intensity value at each sampling time; Indicates the first The number of photons detected in each detection area within the sampling time window; 1 indicates a single sampling time window.

[0045] No. The photon count detected in each detection area within the sampling time window is obtained by spatially partitioning the separation channel using a photoelectric detection array. First, the imaging field of view is divided into multiple detection areas based on the length and height directions of the separation channel, and a region number is assigned to each detection area. The detection area is one of the fixed spatial detection units. During the application of low-intensity periodic perturbations, the photoelectric detection array uses gated sampling to detect the first... The system continuously detects each detection area, converting the weak light signal generated by the bioluminescence of Noctiluca scintillans within that area into electrical pulse signals. A counting circuit or image sensor then reads the number of valid pulses exceeding the background noise threshold within the corresponding sampling time window, thus obtaining the first pulse. The number of photons detected in each detection area within the sampling time window.

[0046] The single sampling time window is determined based on the frequency of the low-intensity periodic perturbation, the duration of Noctiluca scintillans luminescence, and the response speed of the photoelectric detection array. First, the perturbation period is determined based on the perturbation frequency, and then one perturbation period is divided into multiple sampling time windows so that multiple sets of luminescence intensity values ​​can be obtained within each perturbation period, thereby ensuring that a time-series curve of luminescence intensity changing with time can be formed subsequently. If the single sampling time window is too long, it will weaken the temporal resolution of the luminescence peak; if it is too short, it will lead to too few photon counts and a decrease in the signal-to-noise ratio. Through preliminary experiments, the clarity of the luminescence peak and the background noise level under different time windows are compared, and the time window that can both identify the peak position and maintain stable counting is selected as the single sampling time window.

[0047] The first is obtained from continuous sampling The time-series curves of luminescence intensity in each detection region are represented as follows:

[0048] ;

[0049] in, Indicates the first The time-series curve of the luminescence intensity of each detection area over time; This indicates the total number of samples taken within a single low-intensity periodic disturbance detection cycle.

[0050] Since Noctiluca scintillans can produce short-term luminescence when subjected to low-intensity periodic perturbations and maintain a recoverable state as long as the perturbation intensity does not exceed the damage threshold, the above-mentioned luminescence intensity time series curve is regarded as a reversible luminescence response. Compared with bubbles, silt particles, and non-luminescent impurities, the luminescence intensity time series curve of the corresponding region of Noctiluca scintillans usually has peak variation characteristics related to the perturbation rhythm.

[0051] The damage threshold was pre-calibrated using the recoverable state of *Noctiluca scintillated* under different perturbation intensities. *Noctiluca scintillated* samples with similar sources, culture conditions, or on-site sampling conditions to the sample to be separated were selected and placed in the calibration channel. Periodic perturbations were applied in ascending order of intensity, and the morphological integrity, luminescence reversibility, buoyancy response recovery, and survival status of the *Noctiluca scintillated* were recorded after each perturbation. A low perturbation intensity was defined as one where the *Noctiluca scintillated* maintained its cell outline intact, showed no leakage of contents, rupture or collapse, or continued adhesion and inactivation after perturbation, and its peak luminescence intensity recovered to the normal response range after the perturbation ceased. At the damage threshold; when the perturbation intensity continues to increase, and the Noctiluca scintillans exhibits irreversible deformation, significant attenuation or disappearance of luminescence response, loss of periodicity in buoyancy response, cell membrane rupture, and a significant decrease in individual integrity, the shear condition, pressure fluctuation condition, or driving force condition corresponding to the perturbation intensity is determined to be entering the damage range; to ensure control safety, the perturbation intensity at the first occurrence of damage is not directly used as the damage threshold, but rather the maximum safe value corresponding to the first level of perturbation condition before the first occurrence of irreversible damage is selected, or the conservative lower limit value when the integrity rate remains above the preset requirement in multiple repeated calibrations is selected as the final damage threshold.

[0052] S13, while acquiring the reversible luminescence response, the particle image velocimetry unit arranged on the side of the separation channel is activated to continuously image the particles and individual Noctiluca scintillans in the separation liquid and record their displacement changes in the vertical direction; the particle image velocimetry unit is not simply used to determine the presence of particles but to acquire the vertical displacement, instantaneous floating and sinking speed and speed change period of the particles under low-intensity periodic disturbance.

[0053] For the The vertical displacement of a tracked target at adjacent sampling times is expressed as:

[0054] ;

[0055] in, Indicates the first The tracked target was in the first Vertical displacement within each sampling interval; Indicates the first The tracked target was in the first Vertical position at each sampling time; Indicates the first The tracked target was in the first The vertical position at the sampling time is obtained by the particle image velocimetry unit after performing target recognition, coordinate conversion, and trajectory association on continuous image frames. At each sampling time, the particle image velocimetry unit acquires an image frame of the liquid to be separated from the side of the separation channel. Based on a preset grayscale threshold, edge contour, or target brightness difference, it identifies particles and individual Noctiluca scintillans in the image frame. Then, it extracts the contour region of each identified target and calculates the geometric center or brightness weighted center of the contour region. This center point is used as the pixel position of the target in the current image frame. Since the particle image velocimetry unit images from the side of the separation channel, the vertical pixel coordinates in the image coordinate system correspond to the vertical position within the separation channel.

[0056] No. The instantaneous buoyancy velocity of each tracked target is expressed as:

[0057] ;

[0058] in, Indicates the first The tracked target was in the first Instantaneous buoyancy velocity within each sampling interval; This indicates the vertical displacement within the sampling interval; 2 indicates the time interval between two adjacent samples. The time interval between two adjacent samples is obtained by the sampling trigger clock of the particle image velocimetry unit. After the separation device starts low-intensity periodic disturbance, the controller sends a synchronous sampling trigger signal to the particle image velocimetry unit, so that the particle image velocimetry unit continuously acquires the side image of the separation channel according to the preset sampling frequency. When each frame of image is acquired, the controller or image acquisition module synchronously records the timestamp of the frame of image. The difference between the timestamps corresponding to two adjacent frames of image is the time interval between two adjacent samples.

[0059] During continuous disturbance, based on the change of instantaneous buoyancy velocity with time, the first... The velocity change sequence of each tracked target is represented as:

[0060] ;

[0061] in, Indicates the first A sequence of instantaneous buoyancy velocity of a tracked target; This indicates the total number of samples taken.

[0062] The velocity change period is determined based on the time interval between adjacent peak values ​​in the same direction or adjacent zero-crossing points in the instantaneous buoyancy velocity sequence:

[0063] ;

[0064] in, Indicates the first The velocity change cycle of the tracked target; Indicates the first The first tracked target The moment corresponding to the peak velocity in the same direction or the zero crossing point in the same direction; Indicates the first The first tracked target The moment corresponding to the peak velocity or zero-crossing point in the same direction, the first... The first tracked target The moment corresponding to the peak velocity or zero-crossing point in the same direction is obtained from the instantaneous buoyancy velocity sequence analysis of the tracked target. After obtaining the instantaneous floating and sinking velocity sequence of the tracked target, the velocity sequence is first smoothed to eliminate velocity jumps caused by image positioning errors, short-term occlusion, and flow field disturbances. Then, local extrema and zero-crossing points in the velocity sequence are detected in chronological order. The velocity change period is determined using unidirectional velocity peaks. First, the velocity direction is determined, such as selecting upward floating velocity peaks or downward sinking velocity peaks as unidirectional peak types. Then, the velocity sequence is searched for upward peaks where the preceding and following velocity values ​​are both lower than the velocity value at that point, or downward peaks where the preceding and following velocity values ​​are both higher than the velocity value at that point. The 1st... The sampling time corresponding to the local peak that satisfies the same direction condition is denoted as the i-th. The first tracked target The moment corresponding to the peak velocity in the same direction.

[0065] The slow floating and sinking response is composed of vertical displacement, instantaneous floating and sinking velocity, and velocity change period. Because intact Noctiluca scintillans has a large cell volume, weak active movement ability, and relatively slow floating and sinking characteristics, its slow floating and sinking response is significantly different from the rapid buoyancy of bubbles, the settling of mud and sand particles, and the irregular drifting of debris. Therefore, it serves as a basis for identifying the Noctiluca scintillans response separation zone.

[0066] S14, after obtaining the reversible light emission response and the slow floating and sinking response, the time-series curve of the reversible light emission response is spatiotemporally coupled and matched with the velocity change period of the slow floating and sinking response.

[0067] First, the perturbation phase of the periodic driving force at any given moment is defined as follows:

[0068] ;

[0069] in, Indicates the periodic driving force at time 10:00. The perturbation phase; Indicates the frequency of the periodic driving force; Indicates the duration of the disturbance; Indicates the initial phase.

[0070] For the The time series curves of luminescence intensity for each detection region are extracted, and the time corresponding to the peak luminescence intensity is represented as follows:

[0071] ;

[0072] in, Indicates the first The detection area in the first Peak luminous intensity moments within a single peak search window; Indicates the first Each detection area at time The luminous intensity value; Indicates the first A peak search window, at the frequency of the perturbation unit. After outputting the periodic driving force, the controller first determines a single disturbance period based on the disturbance frequency. Then, using the synchronization trigger time or the initial phase corresponding time of the disturbance unit as the starting point of the first period, the continuous disturbance time is sequentially divided into multiple periodic windows, among which the periodic window corresponds to the periodic driving force. The time range corresponding to the first disturbance period can be used as the first... A peak search window.

[0073] The phase difference of the luminous response, obtained from the peak time of luminous intensity and the perturbation phase, is expressed as:

[0074] ;

[0075] in, Indicates the first The detection area in the first Phase difference in luminescence response at each peak; This indicates the perturbation phase corresponding to the peak value of the luminous intensity; Indicates the first The reference phase for each disturbance cycle.

[0076] When the same detection area meets the following condition within multiple consecutive perturbation cycles, the peak emission intensity of the detection area is determined to be constant, expressed as:

[0077] ;

[0078] in, Indicates the calculation of standard deviation; Indicates the first Each detection area in continuous Phase difference in luminescence response within one perturbation cycle; The allowable phase difference fluctuation threshold is obtained through equipment pre-calibration and sample verification. Before formal separation, samples of Noctiluca scintillans with high integrity and clear activity are placed in the separation channel. Multiple rounds of perturbation detection are performed according to the set low-intensity periodic perturbation frequency and amplitude. The phase difference of the luminescence response in the region where the intact Noctiluca scintillans is located is recorded in multiple perturbation cycles, and its phase difference standard deviation is calculated to obtain the phase fluctuation range corresponding to the normal reversible luminescence response of the intact Noctiluca scintillans. At the same time, the same detection is performed on the regions of bubbles, non-luminescent impurities, damaged Noctiluca scintillans, or blank buffer to obtain the phase fluctuation range of non-target regions or abnormal regions. Then, the boundary value that can stably retain the response region of intact Noctiluca scintillans while excluding most abnormal regions is selected as the allowable phase difference fluctuation threshold.

[0079] Standard deviation calculation is used to determine whether the phase difference of the luminescence response in the same detection area is stable over multiple consecutive perturbation cycles. First, the first standard deviation is obtained. Each detection area in continuous The phase difference of the luminous response within each disturbance cycle is denoted as follows: Then, the average value of the multiple light emission response phase differences is calculated to characterize the average phase shift of the detection area during continuous disturbance. Next, the deviation between each light emission response phase difference and the average value is calculated, and each deviation is squared to obtain the phase difference dispersion. Then, all the squared deviations are summed and averaged. Finally, the square root of the average result is taken to obtain the standard deviation of the phase difference of the detection area.

[0080] Simultaneously, based on the slow buoyancy response obtained from the particle image velocimetry unit, the buoyancy velocity amplitude of the tracked target within the corresponding detection area is calculated and expressed as follows:

[0081] ;

[0082] in, Indicates the first The amplitude of the buoyancy velocity of the tracked target within the detection area; Indicates the first The minimum vertical velocity of the tracked target within a detection area during the detection cycle; Indicates the first The maximum vertical velocity of the tracked target within the detection area during the detection period is obtained by filtering the instantaneous buoyancy velocity sequence of all tracked targets within the detection area. The particle image velocimetry unit first selects the maximum vertical velocity of the tracked target within the detection area during the detection period. The Noctiluca scintillans or particulate matter within each detection area are continuously tracked to obtain the vertical positional changes of each tracked target within the detection period. Instantaneous buoyancy is calculated based on the vertical displacement at adjacent sampling times, and then the target is classified into several categories. The instantaneous buoyancy and sinking velocities of the same or multiple tracked targets in each detection area are arranged in chronological order to form a velocity sequence, and abnormal velocity values ​​caused by occlusion, recognition jumps, or trajectory errors are removed. After smoothing and outlier removal, if only the buoyancy and sinking response of a representative target in the area is evaluated, the maximum positive vertical velocity of that representative target within the detection period is selected as the maximum vertical velocity. If the first... If multiple targets exist within a detection area, the maximum vertical velocity of each target can be obtained first. Then, the target with the highest overlap with the luminescence response position, the contour size conforming to the characteristics of Noctiluca scintillans, and the best trajectory continuity can be selected as the effective target, and its maximum vertical velocity can be used as the metric. .

[0083] Determine whether the amplitude of the buoyancy velocity falls within the preset range of algal living characteristics:

[0084] ;

[0085] in, This indicates the lower limit of the velocity amplitude within the preset characteristic range of living algae; This indicates the upper limit of the velocity amplitude within the preset characteristic range of living algae; Indicates the first The amplitude of the floating and sinking speed of the tracked target within the detection area.

[0086] The upper and lower limits of the velocity amplitude within the pre-defined characteristic range of live algae were obtained by calibrating the buoyancy and sinking responses of intact *Noctiluca scintillans* samples under low-intensity periodic perturbation. Individual *Noctiluca scintillans* of different sizes and active states, all remaining intact, were selected as calibration samples. Detection was performed under the same separation channel, buffer conditions, perturbation frequency, and perturbation amplitude as the formal separation. The vertical velocity changes of each intact *Noctiluca scintillans* individual were recorded over multiple perturbation cycles, and the corresponding buoyancy and sinking velocity amplitudes were calculated. Simultaneously, the same tests were performed on non-target or abnormal targets such as bubbles, sediment particles, algal fragments, and damaged *Noctiluca scintillans* to obtain their velocity amplitude distributions. Intact *Noctiluca scintillans* typically exhibits relatively gentle velocity amplitudes with a periodic correspondence; bubbles usually have larger velocity amplitudes; the velocity amplitudes of settled particles may deviate from periodic changes; and the velocity amplitudes of damaged *Noctiluca scintillans* may be lower or unstable. Therefore, the conservative lower value in the velocity amplitude distribution of intact *Noctiluca scintillans* samples was used as the lower limit of the pre-defined characteristic range of live algae velocity amplitude, and the conservative higher value was used as the upper limit of the velocity amplitude.

[0087] When a continuous liquid layer region simultaneously satisfies two conditions—a constant peak luminescence intensity and a constant perturbation phase difference, and a floating / sinking velocity amplitude within a preset characteristic range for living algae—this continuous liquid layer region is determined as a candidate response region. The determination relationship is expressed as follows:

[0088] ;

[0089] in, Indicates the first Whether a detection region belongs to a candidate response region; when When, it indicates the first Each detection region belongs to a candidate response region; when When, it indicates the first The detected area does not belong to the candidate response area; This indicates the permissible phase difference fluctuation threshold; and These represent the lower and upper limits of the velocity amplitude within the preset characteristic range of living algae, respectively.

[0090] Finally, spatial connectivity is determined for all candidate response regions that meet the conditions. The liquid layer region that is continuously distributed within the separation channel and can maintain a stable position update with low-intensity periodic disturbances is dynamically labeled as the Noctiluca scintillans response separation zone, represented as follows:

[0091] ;

[0092] in, Indicates time The separation band of Noctiluca scintillation response obtained by dynamic calibration; This indicates a spatial connectivity filtering operation; Indicates the first Each detection area belongs to a candidate response area; Indicates the first The spatial position of each detection area within the separation channel is obtained from the spatial calibration results of the photoelectric detection array and the particle image velocimetry unit. First, a channel coordinate system is established using the bottom baseline, entrance baseline, or channel centerline of the separation channel. The separation channel is then divided into multiple fixed detection areas along its length and height, each corresponding to a spatial number. Subsequently, a mapping relationship between the pixel positions of the photoelectric detection array, the image positions of the particle image velocimetry unit, and the actual spatial coordinates of the separation channel is established using a calibration plate, standard fluorescent particles, or known geometric markers within the channel. During formal detection, the first... Spatial location of each detection area The detection area can be represented by the coordinates of its center point, boundary coordinates, or the height range of the liquid layer. For example, it can be represented as the length and vertical height within the separation channel. If the separation channel is a two-dimensional observation structure, then... This represents the planar position of the detection area in the channel coordinate system; if the separation channel needs to consider the thickness direction, then... It further includes the depth direction location or is represented by multiple depth layer locations.

[0093] The spatial connectivity region filtering operation involves spatially organizing the detected regions that have been identified as candidate response regions to eliminate isolated noise points and scattered false positive regions. First, based on the aforementioned judgment results, all regions that meet the criteria are read. The detection areas are defined and projected onto the channel coordinate plane according to their spatial positions in the separation channel. Then, it is determined whether adjacent detection areas are connected in the length or height direction, adjacent, or the distance between them is less than the preset connection distance. If two candidate response areas can be connected by consecutive adjacent candidate response areas, they are classified into the same connected area. Then, the area, liquid layer thickness, length extension range, and positional continuity of each connected area are calculated. Isolated areas that are too small, formed by only a single detection area, have abrupt changes in liquid layer thickness, or have a large positional offset from the previous moment are eliminated. Finally, from the remaining connected areas, connected liquid layer areas that are continuously distributed along the flow direction of the separation channel, have a gentle change in vertical height, and can maintain stable position updates during continuous disturbance cycles are selected as the Noctiluca scintillans response separation zone.

[0094] The separation zone of the Noctiluca scintillans response is not determined directly based on the target shape or luminescence brightness, but is determined by simultaneously utilizing the coupling relationship between the reversible luminescence response and the slow floating and sinking response; thereby eliminating brightness disturbances caused by rapid bubble rising, velocity interference caused by non-luminescent impurities, and abnormal luminescence fluctuations caused by damaged Noctiluca scintillans.

[0095] S2, the Noctiluca scintillans response separation band is input into the neural network control model, which distinguishes between intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles and non-luminescent impurities, and generates a response-maintaining separation control command; the response-maintaining separation control command is used to determine the disturbance pause time, the separation inlet alignment position, the buffer escort flow rate and the collection valve opening sequence;

[0096] The neural network control model is formed by concatenating a spatiotemporal feature extraction backbone network, an encoding / decoding semantic segmentation network, a morphology preservation evaluation branch, and a control command generation branch. The spatiotemporal feature extraction backbone network receives a three-dimensional spatiotemporal feature tensor, which originates from multiple frames of two-dimensional spatial distribution images or light intensity-displacement coupling data of the Noctiluca scintillans response separation band within a continuous time window. Therefore, the model input is not an indiscriminate image of the entire separation channel, but rather the calibrated Noctiluca scintillans response separation band. The spatiotemporal feature extraction backbone network employs a structure combining three-dimensional convolutional layers or two-dimensional convolutional layers with temporal units to extract changes in luminescence intensity, vertical displacement, target contour changes, and motion continuity between adjacent frames within the Noctiluca scintillans response separation band. The encoding / decoding semantic segmentation network... The code semantic segmentation network is used to map the above spatiotemporal features back to the spatial location of the separation channel, and output the classification confidence of intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles and non-luminescent impurities on a pixel-by-pixel or per-connected-domain basis. The morphology preservation evaluation branch is connected after the encoded features and segmentation results, and is used to output the morphology preservation coefficient based on the contour continuity, area change stability and boundary contraction degree of the continuous region where intact Noctiluca scintillans is located. The control command generation branch then outputs response preservation separation control commands based on the distribution centroid of intact Noctiluca scintillans, the location of the density maximum, the trend of the morphology preservation coefficient, the nearest distance of the interfering object, the relative motion trend and the distance from the collection port, including the separation inlet alignment position, the disturbance pause time, the buffer escort flow rate and the collection valve opening sequence.

[0097] The neural network control model is constructed according to the detection and execution link of the separation device. First, a unified coordinate system for the separation channel is established based on the separation channel size, photoelectric detection array arrangement, particle image velocimetry unit sampling method, and collection port location. This ensures that the light intensity, displacement, velocity data, and target spatial position in the Noctiluca scintillans response separation zone are mapped to the same input coordinate system. Then, the obtained Noctiluca scintillans response separation zone is used as the data extraction area to construct multi-channel input data. Each data channel corresponds to luminous intensity, vertical displacement, instantaneous buoyancy velocity, response separation zone mask, and historical frame difference information, respectively. Finally, the model output is designed. The semantic segmentation output corresponds to four types of targets: intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles, and non-luminescent impurities. The morphology preservation output corresponds to the morphology preservation coefficient of intact Noctiluca scintillans, and the control output corresponds to the actual executable pump, valve, and inlet control quantities of the separation equipment. To ensure that the output of the neural network control model can be directly used for equipment control, the movable range of the separation inlet, the safe flow range of the buffer pump, the pause response time of the disturbance unit, and the opening and closing delay of the collection valve need to be written into the post-processing rules of the model during construction. This ensures that the response-preserving separation control commands generated by the neural network control model always fall within the control range that the separation equipment can execute without damaging the Noctiluca scintillans.

[0098] The neural network control model is trained through labeled sample training, equipment control sample calibration, and online feedback fine-tuning. During the labeled sample training phase, data on the separation process of *Noctiluca scintillans* from different batches, densities, and backgrounds with different impurities are collected. The locations and categories of intact *Noctiluca scintillans*, damaged *Noctiluca scintillans*, bubbles, and non-luminescent impurities are manually observed or microscopically verified. Simultaneously, the continuity of the outline of intact *Noctiluca scintillans*, the stability of its area, and the presence of adherent aggregation or morphological collapse are recorded to train the semantic segmentation network and the morphology preservation evaluation branch. During the equipment control sample calibration phase, the perturbation pause times, the alignment position of the separation inlet, and the buffer settings (obtained by expert settings or experimental optimization) are recorded on the actual separation equipment. The flow rate of the liquid escort and the opening sequence of the collection valve, combined with the purity and integrity of the separated *Noctiluca scintillans* solution, were used to screen out control samples that could obtain high purity and high integrity, which were then used to train the control command generation branch. In the online feedback fine-tuning stage, the changes in the purity and integrity of the obtained *Noctiluca scintillans* separation solution were used as the separation feedback state to correct the output of the neural network control model. When the impurity contamination increased, the model's sensitivity to the approach trend of bubbles and non-luminescent impurities was increased. When the integrity decreased, the constraint weight of the decreasing trend of the morphology retention coefficient on the disturbance pause time and the buffer escort flow rate was increased, so that the neural network control model could gradually adapt to the differences in the state of *Noctiluca scintillans* in the solution to be separated from different sources.

[0099] S21, after obtaining the Noctiluca scintillans response separation band, the controller continuously samples the Noctiluca scintillans response separation band within a continuous time window to obtain multiple frames of two-dimensional spatial distribution images or light intensity-displacement coupling data; the two-dimensional spatial distribution images are used to characterize the spatial distribution state of the Noctiluca scintillans response separation band in the separation channel; the light intensity-displacement coupling data are used to simultaneously characterize the reversible luminescence response and the slow buoyancy response, enabling the neural network control model to identify the target category and response stability in the same input data.

[0100] After dynamically calibrating and obtaining the Noctiluca scintillans response separation zone, the controller first reads the boundary position, central liquid layer height, and time-updated spatial range of the Noctiluca scintillans response separation zone in the separation channel coordinate system. This spatial range is then set as the target interception area for continuous sampling. Subsequently, the controller triggers the photoelectric detection array and particle image velocimetry unit to perform multiple samplings within a preset continuous time window using a unified synchronous clock. At each sampling moment, the photoelectric detection array acquires the luminous intensity or photon count distribution of each detection point within the target interception area, forming a two-dimensional spatial distribution image reflecting the spatial brightness state of the Noctiluca scintillans response separation zone. Simultaneously, the particle image velocimetry unit acquires the same... The system captures the vertical positional changes of Noctiluca scintillans, bubbles, and impurity particles within a target area. The vertical displacement and instantaneous buoyancy velocity are obtained from the positional changes at adjacent sampling times. Then, based on the spatial calibration relationship between the photoelectric detection array and the particle image velocimetry unit, the controller aligns the luminous intensity value, vertical displacement, instantaneous buoyancy velocity, and Noctiluca scintillans response separation band marker value at the same spatial location to the same coordinate grid, forming the light intensity-displacement coupled data at that sampling time. After multiple samplings are completed within a continuous time window, the controller can obtain multiple frames of two-dimensional spatial distribution images and multiple frames of light intensity-displacement coupled data arranged in the acquisition sequence.

[0101] In the At each sampling time, the input frame corresponding to the Noctiluca scintillans response separation band is represented as:

[0102] ;

[0103] in, Indicates the first At each sampling time in spatial location Place, No. Input values ​​for each data channel; Indicates the position along the length direction in the coordinate system of the separation channel; Indicates the position in the height direction within the separation channel coordinate system; The data channel number is indicated. When the input data is a two-dimensional spatial distribution image, the data channel is the gray value of the response area or the luminous intensity value. When the input data is light intensity-displacement coupled data, the data channel corresponds to the luminous intensity value, vertical displacement, instantaneous buoyancy velocity, or response separation zone marker value, respectively.

[0104] Collected within a continuous time window The input frames are stacked according to the acquisition time sequence to obtain the three-dimensional spatiotemporal feature tensor representation as follows: ;in, Represents a three-dimensional spatiotemporal feature tensor; This indicates a stacking operation performed in chronological order. Indicates the first Two-dimensional spatial distribution image or light intensity-displacement coupled data corresponding to each sampling time; This indicates the number of frames collected within a continuous time window.

[0105] To ensure consistent data scale across different times and detection areas, each input frame is normalized before stacking, as follows:

[0106] ;

[0107] in, This represents the normalized input value; This represents the input value before normalization; Indicates the first The average of the data channels; Indicates the first Standard deviation of each data channel; This represents the smallest positive number that prevents the denominator from being zero.

[0108] The normalized consecutive input frames are represented according to the acquisition time sequence as follows: ;in, This represents the normalized three-dimensional spatiotemporal feature tensor.

[0109] S22, after inputting the three-dimensional spatiotemporal feature tensor into the neural network control model, the neural network control model first extracts the spatial distribution changes, luminescence intensity changes, and buoyancy displacement changes of the Noctiluca scintillans response separation zone within a continuous time window through a spatiotemporal feature extraction layer, and obtains the spatiotemporal feature encoding result as follows:

[0110] ;

[0111] in, This represents the spatiotemporal feature encoding result; This represents the spatiotemporal feature extraction function in the neural network control model; This represents the normalized three-dimensional spatiotemporal feature tensor.

[0112] The spatiotemporal feature extraction function in the neural network control model consists of a three-dimensional convolutional feature extraction layer, a temporal attention layer, and a multi-scale fusion layer. First, the normalized three-dimensional spatiotemporal feature tensor is input into the three-dimensional convolutional feature extraction layer. The three-dimensional convolutional kernel slides simultaneously along the time, length, and height directions to extract the changes in luminescence intensity, vertical displacement, instantaneous floating and sinking velocity, and boundary morphology of the Noctiluca scintillans response separation zone within a continuous time window. Then, the features obtained from the three-dimensional convolution are input into the temporal attention layer. The temporal attention layer compares the response continuity between adjacent frames, enhances features that change synchronously with low-intensity periodic disturbances, and suppresses interference features caused by short-term noise, bubble flickering, and random displacement of non-luminescent impurities. Finally, the local contour features, regional motion features, and overall liquid layer change features output from the convolutional layers at different scales are spliced ​​and fused to obtain the spatiotemporal feature encoding result.

[0113] The neural network control model performs encoding-decoding semantic segmentation on the spatiotemporal feature encoding results, and obtains the classification score of each category at each pixel position or each connected component position as follows: ;in, This represents the set of classification scores output by semantic segmentation. Represents the encoding-decoding semantic segmentation function; This represents the spatiotemporal feature encoding result.

[0114] The encoder-decoder semantic segmentation function consists of a downsampling encoder, a skip connection layer, and an upsampling decoder. The encoder first downsamples the spatiotemporal feature encoding results step by step to gradually expand the receptive range, enabling the neural network control model to identify the overall distribution relationship of intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles, and non-luminescent impurities in the Noctiluca scintillans response separation band. During the downsampling process, shallow features mainly retain the target edge, luminescent boundary, and local displacement differences, while deep features mainly retain the target category, morphological stability, and spatial relationship with surrounding interference objects. Subsequently, the shallow edge features are passed to the corresponding scale decoding stage through the skip connection layer to avoid the loss of Noctiluca scintillans contour edges after multiple downsampling. Finally, the decoder upsamples step by step to restore the low-resolution semantic features to the original spatial coordinate scale and outputs the classification scores of intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles, and non-luminescent impurities at each pixel position or each connected component position. The classification confidence is then obtained after normalization.

[0115] Normalizing the classification scores yields the classification confidence score, expressed as follows: ;in, Indicates spatial location Belongs to the Classification confidence of the target class; Indicates spatial location Belongs to the Classification score of the target category; This indicates the total number of target categories, which include intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles, and non-luminescent impurities.

[0116] For each spatial location, the target category is determined based on the maximum classification confidence score, represented as follows: ;in, Indicates spatial location The classification results; This indicates that the category with the highest classification confidence is selected from all target categories; Indicates spatial location Belongs to the Classification confidence of target classes.

[0117] When using connected component-by-connection classification, the first step is to obtain the first connected component based on the semantic segmentation results. Each connected component is represented as:

[0118] ;

[0119] in, Indicates the first One connected component; Indicates spatial location within a connected domain; Indicates the first The target category corresponding to each connected component.

[0120] No. The target category corresponding to each connected component is determined by the classification results and classification confidence of each spatial location within that connected component. First, based on the pixel-level classification results output by the encoder-decoder semantic segmentation function, spatially adjacent pixels or sampling points with initially consistent categories are merged to obtain the target category of the first connected component. The system calculates the classification confidence scores of all pixels within a connected component, identifying them as belonging to intact Noctiluca scintillans, damaged Noctiluca scintillans, air bubbles, and non-luminescent impurities. It also calculates the average classification confidence score percentage for each category within the connected component. If a category has the highest average classification confidence score and exceeds a preset category confirmation threshold, then that category is designated as the nth connected component. The target category corresponding to each connected component.

[0121] The category confirmation threshold is obtained through statistical analysis of validation samples from the semantic segmentation model. First, separation process data are collected under different water sample concentrations, different Noctiluca scintillans activity states, and different conditions of bubble and impurity interference. The true category of each connected component is determined by manual microscopic verification or expert annotation, forming a validation sample set of intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles, and non-luminescent impurities. Then, the validation sample set is input into the trained neural network control model. The highest classification confidence score corresponding to each connected component being correctly classified and the highest classification confidence score corresponding to being misclassified or confused are statistically analyzed. The confidence score distributions of correctly classified samples and misclassified samples are then compared, and the confidence score cutoff value that simultaneously meets the preset requirements for the accuracy of intact Noctiluca scintillans identification, impurity removal rate, and damaged individual removal rate is selected as the category confirmation threshold.

[0122] No. The classification confidence of each connected component is represented as follows:

[0123] ;

[0124] in, Indicates the first Classification confidence of each connected component; Indicates the first The number of pixels or spatial sampling points within a connected region; Indicates spatial location Category The classification confidence level.

[0125] time The continuous region classified as a complete Noctiluca scintillans by the neural network control model To determine whether the intact *Noctiluca scintillans* remains in a suitable state for separation, the neural network control model further outputs a morphology retention coefficient. This coefficient characterizes the outline integrity and area stability of the continuous region containing the intact *Noctiluca scintillans*. The area of ​​the continuous region containing the intact *Noctiluca scintillans* is represented as:

[0126] ;

[0127] in, Indicates time The area of ​​the continuous region containing the complete Noctiluca scintillans; This indicates the number of pixels or spatial sampling points within a complete continuous area of ​​Noctiluca scintillans.

[0128] The integrity of the outline of the continuous region containing the complete Noctiluca scintillans is represented as follows:

[0129] ;

[0130] in, Indicates time The integrity of the outline of the continuous region containing the complete Noctiluca scintillans; This represents the area of ​​the continuous region containing the complete Noctiluca scintillans algae; This represents the perimeter of the continuous region containing the complete Noctiluca scintillans. The perimeter of the continuous region containing the complete Noctiluca scintillans is calculated from the boundary of the region obtained by semantic segmentation. First, continuous regions classified as complete Noctiluca scintillans are selected from the classification results, and the boundary of the continuous region is extracted to determine its outer contour pixels or outer contour sampling points. Then, according to the spatial calibration relationship of the separation channels, the pixel spacing in the image is converted into the actual spatial distance to avoid directly substituting the number of pixels for the true perimeter. Then, according to the arrangement order of the contour points, the actual distance between adjacent contour points is calculated in turn, and the distances of all adjacent contour points are accumulated to obtain the perimeter of the continuous region containing the complete Noctiluca scintillans.

[0131] The outline integrity is used to reflect whether the complete Noctiluca scintillans outline is continuous, whether the contraction is obvious, and whether the boundary is broken.

[0132] The area stability of the continuous region containing the complete Noctiluca scintillans is expressed as:

[0133] ;

[0134] in, Indicates time Area stability of the continuous region containing the complete Noctiluca scintillans; This represents the area of ​​the continuous region containing the complete Noctiluca scintillans at the current moment; This represents the average area of ​​the continuous region containing the complete Noctiluca scintillans within a preset time range prior to the current moment; This represents the smallest positive number that prevents the denominator from being zero.

[0135] The shape retention coefficient, derived from contour integrity and area stability, is expressed as:

[0136] ;

[0137] in, Indicates time Morphological preservation coefficient of the continuous region containing the complete Noctiluca scintillans; Indicates the completeness of the outline; Indicates area stability; This represents the weighting coefficient, with a value range of [value range missing]. .

[0138] When the morphology retention coefficient is high, it indicates that the outline of the intact Noctiluca scintillans is continuous and the area changes smoothly, making it suitable for entering the separation control process. When the morphology retention coefficient drops rapidly, it indicates that the intact Noctiluca scintillans may be affected by continuous disturbance, wall adhesion, bubble traction, or negative pressure at the inlet. It is necessary to stop the disturbance and start the escort separation before it drops to the preset threshold.

[0139] The preset threshold usually refers to the preset morphology retention threshold used in judging the decrease in morphology retention coefficient. It is obtained through calibration experiments and separation effect verification. First, select Noctiluca scintillans samples that are in a normal active state and confirmed to be intact by microscopic observation. Apply low-intensity periodic perturbation in the separation device and perform multiple rounds of separation control. Record the morphology retention coefficient of the samples under different states such as undamaged, slightly deformed, obviously shrunken, adherent and aggregated, and broken and inactive. Then, the morphology retention coefficient of the samples that remain intact after separation, whose luminescence response can be recovered, and that can enter the collection branch normally is taken as the safe sample interval. The morphology retention coefficient of the samples that show irreversible deformation, broken outline, abrupt area change, or decreased integrity rate is taken as the risk sample interval. Finally, select the boundary value that can distinguish between the safe sample interval and the risk sample interval as the preset morphology retention threshold.

[0140] S23, the neural network control model generates a response-maintaining separation control command based on the obtained complete Noctiluca scintillans classification results, classification confidence, and morphology retention coefficient; the response-maintaining separation control command is used to determine the separation inlet alignment position, disturbance pause time, buffer escort flow rate, and collection valve opening sequence.

[0141] Based on the centroid and density maxima location of the continuous region containing the intact *Noctiluca scintillans*, the alignment position of the separation inlet is generated; the centroid of the continuous region containing the intact *Noctiluca scintillans* is represented as:

[0142] ;

[0143] in, Indicates time The distribution center of the continuous area where the complete Noctiluca scintillans is located; Indicates the continuous region containing the complete Noctiluca scintillans; Indicates spatial location At any moment The classification confidence level of *Noctiluca scintillans* belonging to the complete *Noctiluca scintillans*; Indicates spatial location Position vector in the separate channel coordinate system, spatial position The position vector in the separation channel coordinate system is obtained through spatial calibration of the separation channel. First, the separation channel inlet end, bottom boundary, or center of the separation suction inlet is used as the coordinate reference to establish the separation channel coordinate system. Indicates the position along the length of the separation channel. The height direction position perpendicular to the bottom of the channel is indicated. Then, using a calibration plate, known channel geometry, standard fluorescent particles, or micro-scale markings, the pixel coordinates of the photoelectric detection array and particle image velocimetry unit are spatially calibrated to obtain the mapping relationship between the pixel coordinates and the actual spatial coordinates of the separation channel. During formal detection, substituting the pixel coordinates corresponding to any spatial position into this mapping relationship yields the actual coordinates of that spatial position in the separation channel coordinate system, and the result is expressed as a position vector. express.

[0144] The location of the density maxima of intact Noctiluca scintillans is represented as follows:

[0145] ;

[0146] in, Indicates time Location of the maximum density of *Noctiluca scintillans*; This indicates selecting the spatial location with the highest spatial density; Indicates time The complete Noctiluca scintillans is located in the position Nearby spatial density, time The complete Noctiluca scintillans is located in the position The spatial density in the vicinity is calculated based on the classification confidence of intact *Noctiluca scintillans* and the spatial distribution of the continuous region where intact *Noctiluca scintillans* is located, as output by the neural network control model. First, pixels, sampling points, or connected component center points classified as intact *Noctiluca scintillans* are selected from the output classification results, and these points are mapped to the separated channel coordinate system. Then, the location to be calculated is... A local neighborhood window is set up around the center. The number, area ratio, or classification confidence weighted sum of intact *Noctiluca scintillans* target points within this neighborhood window are counted. If a confidence weighting method is used, intact *Noctiluca scintillans* points with higher classification confidence contribute more to the spatial density, while points with lower classification confidence or unstable boundaries contribute less. Furthermore, intact *Noctiluca scintillans* points within the neighborhood are ranked according to their distance and location. The distance is weighted, with closer distances having higher weights and farther distances having lower weights, thus obtaining the time. The complete Noctiluca scintillans is located in the position The density of the surrounding space.

[0147] The separation inlet alignment position, generated from the distribution centroid and the location of the density maxima, is represented as follows:

[0148] ;

[0149] in, Indicates time Align the generated separation suction port with the position; This indicates the centroid of the continuous region containing the complete Noctiluca scintillans; This indicates the location of the maximum density of the intact Noctiluca scintillans; This represents the location fusion coefficient, with a value range of [value range missing]. In summary, the separation inlet is not simply aimed at a single bright spot, but at a concentrated and stable area of ​​complete Noctiluca scintillans, thus improving collection stability.

[0150] Based on the rate of change of the shape preservation coefficient over time, a disturbance pause time is generated before the shape preservation coefficient drops to a preset threshold; the rate of change of the shape preservation coefficient over time is expressed as:

[0151] ;

[0152] in, Indicates time The rate of change of the lower morphology retention coefficient over time; This represents the shape preservation coefficient at the current moment; This represents the shape preservation coefficient at the previous shape detection time. The time interval between two consecutive calculations of the morphology preservation coefficient is determined by the morphology evaluation triggering cycle of the neural network control model or controller. After receiving the three-dimensional spatiotemporal feature tensor within the continuous time window, the neural network control model will output the morphology preservation coefficient of the complete Noctiluca scintillans region according to the preset inference frequency. The controller records the timestamp corresponding to each morphology preservation coefficient output. The difference between two consecutive output timestamps is the time interval between two consecutive calculations of the morphology preservation coefficient.

[0153] The preset inference frequency is determined based on the rate of change of the state of Noctiluca scintillans response to the separation zone, the single inference time of the neural network control model, and the response time of the separation device's actuator. First, in the calibration experiment, the shortest change time required for Noctiluca scintillans to change from a stable suspension state to a state of obvious contraction, adhesion aggregation, or risk of breakage under low-intensity periodic disturbance is recorded, and this shortest change time is used as the upper limit of the time for morphological risk identification. Then, the calculation time required for the neural network control model to complete one three-dimensional spatiotemporal feature tensor input, classification confidence output, and morphological retention coefficient output is measured. At the same time, the execution response time of the disturbance unit pausing, the separation inlet alignment, the buffer pump establishing the escort flow rate, and the collection valve opening is measured. The time interval between two adjacent model inferences is made less than the difference between the shortest change time and the execution response time, and a safety margin is reserved to obtain the corresponding preset inference frequency.

[0154] when When the morphology retention coefficient of *Noctiluca scintillans* is decreasing, the predicted time when it decreases to a preset threshold is expressed as:

[0155] ;

[0156] in, This indicates the moment when the predicted shape retention coefficient drops to a preset threshold; Indicates the current time; Indicates the current form retention coefficient; This represents the rate of change of the shape retention coefficient over time. This indicates an extremely small positive number that prevents the denominator from being zero; The preset morphology retention threshold is obtained by calibrating the morphological stability of intact Noctiluca scintillans samples. First, select Noctiluca scintillans samples that are intact and have normal activity through microscopic observation. Then, conduct multiple rounds of separation experiments under the same channel structure, buffer conditions, low-intensity periodic perturbation conditions, and sampling conditions as the formal separation. Record the morphology retention coefficients of the samples under various states, including normal suspension, slight shrinkage, adhesion aggregation, bubble traction, stretching under negative pressure at the inlet, and rupture inactivation. Then, the morphology retention coefficients of samples that still maintain an intact outline, have stable area changes, and recoverable luminescence response after separation are taken as the safe range. The morphology retention coefficients of samples that show irreversible deformation, abrupt area changes, broken outlines, or decreased integrity are taken as the risk range. Finally, select the boundary value that can distinguish between the safe range and the risk range as the preset morphology retention threshold.

[0157] The disturbance pause time generated before the predicted time is represented as follows:

[0158] ;

[0159] in, Indicates the moment when the disturbance pauses; This indicates the moment when the predicted shape retention coefficient drops to a preset threshold; The safety lead time is determined based on the stop response time of the disturbance unit, the movement response time of the separation inlet, the buffer escort flow establishment time, and the rate of morphological instability development of Noctiluca scintillans. During the equipment calibration phase, the time required for the disturbance unit to decrease from receiving the pause command to the actual disturbance amplitude decreasing to a safe range is measured. The time required for the separation inlet to move from its current position to the target alignment position is measured. The time required for the buffer pump to form a stable escort flow rate from receiving the escort command is measured. The time required for the collection valve to reach an effective opening degree from receiving the opening command is also measured. Then, combined with the historical rate of change of the Noctiluca scintillans morphological retention coefficient from its current value to the preset morphological retention threshold, the remaining time available before Noctiluca scintillans enters a morphological risk state is estimated. The safety lead time should be greater than the main delay in the above-mentioned actuator response delay and retain a certain buffer margin to ensure that the disturbance pause, inlet alignment, and buffer escort have already taken effect before the morphological retention coefficient decreases to the preset morphological retention threshold.

[0160] In summary, low-intensity periodic disturbances can be paused before the intact Noctiluca scintillans enters a state of obvious morphological instability, thus avoiding damage or adhesion to the algae due to continuous disturbances.

[0161] Based on the closest distance and relative motion trend between intact Noctiluca scintillans and bubbles and non-luminescent impurities, the flow rate compensation value for generating buffer escort flow rate represents the interference region corresponding to bubbles and non-luminescent impurities as follows:

[0162] ;

[0163] in, Indicates time Lower interference region; Indicates the area where the bubble is located; Indicates the region where non-luminescent impurities are located; This indicates a region merging operation.

[0164] The regions containing bubbles and non-luminescent impurities are determined by the semantic segmentation results and classification confidence scores output by the neural network control model. After performing spatiotemporal feature extraction and encoding-decoding semantic segmentation on the three-dimensional spatiotemporal feature tensor, the neural network control model outputs the classification confidence scores for complete Noctiluca scintillans, damaged Noctiluca scintillans, bubbles, and non-luminescent impurities at each pixel location or each connected component location. When the classification confidence score for a bubble category corresponding to a certain spatial location or connected component is the highest and exceeds the category confirmation threshold, that spatial location or connected component is marked as a bubble region. The region containing bubbles is obtained by merging all adjacent spatial locations or connected components marked as bubbles. Similarly, when the classification confidence score for a non-luminescent impurity category corresponding to a certain spatial location or connected component is the highest and exceeds the corresponding category confirmation threshold, that spatial location or connected component is marked as a non-luminescent impurity region. The region containing non-luminescent impurities is obtained by merging all adjacent spatial locations or connected components marked as non-luminescent impurities.

[0165] The closest distance between the intact Noctiluca scintillans and the disturbed area is represented as:

[0166] ;

[0167] in, Indicates time The closest distance between the intact Noctiluca scintillans and the disturbed area; Indicates the spatial location of the complete Noctiluca scintillans within a continuous region; Indicates the spatial location within the interference area; This represents the Euclidean distance.

[0168] The relative motion trend between the intact Noctiluca scintillans and the disturbed area is represented as follows:

[0169] ;

[0170] in, This indicates the relative movement trend between the intact Noctiluca scintillans and the disturbed area; Indicates the location of the nearest intact Noctiluca scintillans to the interference area; Indicates the location of the closest disturbing area to the intact Noctiluca scintillans; This indicates the movement speed of the intact Noctiluca scintillans; Represents the velocity of the interference area; · represents the vector dot product; if , indicating that the interfering substances tend to be close to the intact Noctiluca scintillans; if 0 indicates that the interfering substances do not show a trend of approaching the intact Noctiluca scintillans.

[0171] The flow rate compensation value for buffer escort flow rate is expressed as:

[0172] ;

[0173] in, Indicates time The flow rate compensation value for the buffer escort flow rate; Indicates the distance compensation coefficient; Indicates the preset safe distance; This indicates the closest distance between the intact Noctiluca scintillans and the disturbed area; Indicates the relative motion compensation coefficient; This indicates the relative movement trend between the intact Noctiluca scintillans and the disturbed area.

[0174] The distance compensation coefficient, preset safety distance, and relative motion compensation coefficient were obtained through separation equipment calibration experiments and separation effect feedback. First, different initial distances and relative motion states were set between intact *Noctiluca scintillans* and bubbles / non-luminescent impurities in standard or simulated water samples. Different intensities of buffer escort flow rates were applied, and the following were recorded: whether the interfering substances were pushed away, whether the intact *Noctiluca scintillans* maintained its original response separation zone, and whether it adhered to the wall, deformed, or was dragged by impurities. Then, the minimum distance that could stably prevent bubbles or non-luminescent impurities from approaching the intact *Noctiluca scintillans* without damaging it was statistically determined, and this distance was defined as the preset safety distance. The distance compensation coefficient represents the proportion by which the buffer escort flow rate should increase with the degree of distance insufficiency when the actual closest distance is less than the preset safety distance. This was achieved by fitting experimental data of multiple sets of escort flow rate compensation values ​​required for insufficient distance. The slope that can push away the interfering material without significantly reducing the morphological preservation coefficient of *Noctiluca scintillans* was obtained and selected as the distance compensation coefficient. The relative motion compensation coefficient was obtained by calibrating the escort flow rate during the approach of *Noctiluca scintillans* to the interfering material. First, different types of interfering materials, including bubbles, non-luminescent impurities, and debris particles, were set in the separation channel and brought close to the continuous area where *Noctiluca scintillans* was located at different relative approach velocities. Then, different sizes of buffer escort flow rates were applied, and the minimum flow rate compensation value that could push the interfering material away from *Noctiluca scintillans* without significantly reducing the morphological preservation coefficient of *Noctiluca scintillans* at each relative approach velocity was recorded. Then, the relative approach trend intensity and the required minimum flow rate compensation value were statistically correlated to obtain the relationship between the two, and the effective slope or conservative fitting coefficient in this relationship was selected as the relative motion compensation coefficient.

[0175] When the interfering object is close to or has a tendency to approach the intact Noctiluca scintillans, the flow rate compensation value increases, so that the buffer solution can push the interfering object away during the escort process; when the interfering object is far away and has no tendency to approach, the flow rate compensation value decreases, so as to avoid the morphological instability of the intact Noctiluca scintillans caused by excessive flow rate.

[0176] Based on the current longitudinal displacement velocity and distance from the collection port of the intact *Noctiluca scintillans*, the predicted time window for its arrival at the collection port is calculated, and the opening sequence of the collection valve is generated based on the predicted time window; the distance from the intact *Noctiluca scintillans* to the collection port is expressed as:

[0177] ;

[0178] in, Indicates time Distance between the complete Noctiluca scintillans and the collection port; Indicates the alignment position of the separation suction port; This indicates the spatial location of the collection port in the separation channel coordinate system.

[0179] The predicted arrival time of the intact *Noctiluca scintillans* at the collection port is expressed as:

[0180] ;

[0181] in, This indicates the predicted arrival time of the complete Noctiluca scintillans at the collection port; Indicates the current time; This indicates the distance of the complete Noctiluca scintillans from the collection port; This indicates an extremely small positive number that prevents the denominator from being zero; The current longitudinal displacement velocity of the intact *Noctiluca scintillans* is determined jointly by the obtained slow buoyancy response and the obtained classification results of the intact *Noctiluca scintillans*. The time is first determined based on the semantic segmentation results of the neural network control model. The continuous region containing the complete Noctiluca scintillans is selected, and the main target or representative connected region corresponding to the alignment position of the separation inlet is selected within this continuous region. Then, the longitudinal position of the main target at the current sampling time and the previous sampling time is read. The longitudinal position can be obtained by spatial calibration conversion of the image coordinates of the particle image velocimetry unit. The difference between the current longitudinal position and the previous longitudinal position is then divided by the time interval between two adjacent samplings to obtain the current longitudinal displacement velocity of the complete Noctiluca scintillans.

[0182] The prediction time window is represented as follows:

[0183] ;

[0184] in, This indicates the predicted time window for the complete Noctiluca scintillans to reach the collection port; Indicates the predicted arrival time; This represents the prediction time margin, used to compensate for flow rate disturbances, position recognition errors, and valve response delays. It is determined based on the possible time errors that may occur during the process of intact Noctiluca scintillans reaching the collection port. During the equipment calibration phase, the actual arrival time of intact Noctiluca scintillans from the alignment position of the separation inlet to the collection port is recorded multiple times and compared with the predicted arrival time calculated based on the current longitudinal displacement velocity and the distance from the collection port to obtain the prediction error distribution. Then, the prediction error range is statistically analyzed under different buffer escort flow rates, different Noctiluca scintillans densities, different impurity interference conditions, and different channel positions. Finally, the conservative upper limit of the prediction error is used as the base value of the prediction time margin, and necessary safety margins in the valve opening response time, valve closing lag time, and controller command transmission delay are added to obtain the final prediction time margin.

[0185] The collection valve opening sequence generated based on the predicted time window is expressed as follows:

[0186] ;

[0187] ;

[0188] ;

[0189] in, Indicates the opening sequence of the collection valve; Indicates the moment the collection valve opens; Indicates the time when the collection valve closes; Indicates the predicted arrival time; Indicates the valve body's advance opening time; The valve body delayed closing time is obtained through the collection valve response calibration and the Noctiluca scintillans arrival error statistics. During the separation equipment calibration stage, opening and closing commands are continuously sent to the collection valve. The time required for the collection valve to reach an effective opening state from receiving the opening command is recorded using a valve position sensor, flow sensor, or high-speed imaging. This time is taken as the valve body opening response time. At the same time, the time required for the collection valve to actually cut off or reduce the collection flow to an invalid level from receiving the closing command is recorded. This time is taken as the valve body closing response time.

[0190] The response-holding separation control command ultimately generated by the neural network control model is expressed as follows:

[0191] ;

[0192] in, This indicates a response to a hold-type separation control command; Indicates the alignment position of the separation suction port; Indicates the moment when the disturbance pauses; This indicates the flow rate compensation value for the buffer escort flow rate; This indicates the timing of the collection valve opening.

[0193] The acquired response-maintaining separation control commands are not simply classification results, but directly correspond to the execution actions of the separation equipment. This enables the neural network control model to output control information that can actually drive the separation inlet, disturbance unit, buffer pump, and collection valve based on the distribution of intact Noctiluca scintillans, morphological maintenance status, proximity of interfering substances, and arrival time at the collection port.

[0194] S3, according to the response-maintaining separation control command, the disturbance unit, separation inlet, buffer pump and collection valve are controlled to work together to introduce the complete Noctiluca scintillans into the collection branch while the Noctiluca scintillans response separation zone remains intact and does not adhere to the wall and aggregate, so as to obtain Noctiluca scintillans separation solution.

[0195] S31, after receiving the response-holding separation control command output by the neural network control model, the controller executes the response-holding separation control command... The data is analyzed to extract the disturbance pause time, the alignment position of the separation inlet, the flow rate setpoint of the buffer escort flow, and the opening sequence of the collection valve.

[0196] To ensure that the buffer escort flow rate can simultaneously achieve both longitudinal escort along the separation channel and anti-wall adhesion near the wall, the flow rate setpoint is decomposed into an axial escort component along the separation channel axis and a radial anti-wall adhesion component along the separation channel radially, expressed as follows: ;in, This indicates the flow rate setpoint for buffer escort flow; This indicates the axial escort component along the separation channel axis, used to propel the intact Noctiluca scintillans to move smoothly along the longitudinal direction of the separation channel towards the separation inlet; This represents the radial anti-wall component along the radial direction of the separation channel, used to create a microflow effect near the wall towards the center of the channel.

[0197] The axial escort component along the separation channel axis and the radial anti-adhesion component along the separation channel radial direction are obtained by combining the buffer escort flow rate setpoint output by the neural network control model with the geometric direction of the separation channel, the position of the Noctiluca scintillans response separation zone, and the risk allocation of wall adhesion. The controller first determines the axial and radial directions according to the separation channel coordinate system. The axial direction is the main movement direction of the intact Noctiluca scintillans from the Noctiluca scintillans response separation zone into the separation inlet, and the radial direction is the anti-adhesion direction from the wall of the separation channel to the center of the channel or to the center of the Noctiluca scintillans response separation zone. Then, the distance between the Noctiluca scintillans response separation zone and the separation inlet, the current longitudinal displacement velocity of the intact Noctiluca scintillans, the distance between the Noctiluca scintillans response separation zone and the nearest wall surface, and the trend of the morphology retention coefficient are read.

[0198] S32, when the disturbance pause time is reached, the controller sends a pause command to the disturbance unit, causing the disturbance unit to stop outputting the periodic driving force, expressed as: ;in, Indicates the time of the disturbance element. The periodic driving force output; This indicates the moment when the disturbance pauses.

[0199] After the disturbance unit stops outputting periodic driving force, the periodic microflow caused by low-intensity periodic disturbances in the separation channel gradually decays. The controller simultaneously starts the buffer pump, causing it to establish a laminar escort flow field along the longitudinal direction of the separation channel according to the axial escort component. The axial escort velocity is expressed as: ;in, This represents the mainstream velocity of the laminar escort flow field along the longitudinal direction of the separation channel; This indicates the axial escort component.

[0200] To ensure that the laminar escort flow field does not disrupt the Noctiluca scintillans response separation zone, the axial escort component should satisfy a low shear constraint, expressed as follows: ;in, This represents the shear force exerted by the laminar escort flow field on the Noctiluca scintillans response separation zone; This represents the elastic deformation threshold of the cell membrane of Noctiluca scintillans.

[0201] Driven by the radial anti-adhesion component, the controller controls the buffer pump or lateral microfluidic inlet to form a micro-pressure difference or micro-vortex near the wall of the separation channel towards the center of the channel. This causes the Noctiluca scintillans in the separation zone to be slightly pushed back towards the center, inhibiting its migration and aggregation towards the wall.

[0202] The radial anti-adhesion effect near the wall surface is expressed as: ;in, This indicates the slight pressure difference near the wall towards the center of the channel; Indicates the radial anti-wall component; The conversion factor between radial flow velocity and micro-pressure difference is obtained through fluid calibration experiments in the separation channel. Before the equipment is put into formal use, a calibration solution with viscosity, salinity, and temperature similar to the liquid to be separated is injected into the separation channel, and the separation suction port and collection valve are closed to ensure the separation channel is in a stable liquid-filled state. Then, different radial buffer flow rates are output step by step through the lateral buffer inlet or the radial anti-wall-adhesion inlet. At the same time, a micro-pressure difference sensor is used to detect the micro-pressure difference formed between the area near the wall of the separation channel and the center of the channel, or a particle image velocimetry unit is used to inversely measure the wall pressure. The pressure change corresponding to the radial flow velocity near the surface is measured, and then each level of radial flow velocity is paired with the corresponding measured micro-pressure difference to form radial flow velocity-micro-pressure difference calibration data. In order to ensure that the conversion factor is applicable to the separation process of Noctiluca scintillans, the corresponding micro-pressure difference should also be verified during calibration to ensure that it does not cause Noctiluca scintillans response separation zone breakage, wall rebound, or decrease in morphology retention coefficient. Finally, the calibration range that can form a stable anti-wall adhesion effect near the wall towards the center of the channel and does not exceed the elastic deformation threshold of Noctiluca scintillans cell membrane is selected to determine the conversion factor between radial flow velocity and micro-pressure difference.

[0203] The flow field within the separation channel switches from a periodic disturbance state to a laminar escort state, and the radial anti-attachment component reduces the adhesion, retention, or aggregation of Noctiluca scintillans, providing a stable flow field basis for subsequent separation inlet alignment and suction operations.

[0204] S33, during the alignment and suction operation of the separation suction port, a slight pressure difference is continuously maintained near the wall surface in the direction towards the center of the channel. This ensures that the Noctiluca scintillans response separation zone maintains an anti-adhesion gap with the wall of the separation channel, and uses the anti-adhesion gap as an auxiliary constraint to adjust the preset inhalation rate.

[0205] After the laminar flow escort flow field has stabilized, the controller controls the movement of the separation inlet according to the alignment position of the separation inlet, so that the separation inlet is aligned with the spatial coordinates of the centroid of the complete Noctiluca scintillans distribution; the target position of the separation inlet is represented as: ;in, Indicates the separation of the intake port at time Target alignment position; This indicates the alignment position of the separate intake port generated by the neural network control model.

[0206] To avoid local flow field disturbances caused by excessively rapid movement of the separation inlet, the movement speed of the separation inlet is limited as follows: ;in, Indicates the rate of change of the position of the separation inlet; The maximum allowable moving speed of the separation inlet is determined by the combined effect of the mechanism's response capability and the flow field disturbance limitation. First, the maximum displacement speed, positioning error, and repeatability stability of the separation inlet drive mechanism are tested under no-load conditions to obtain the upper limit of the speed at which the mechanical structure can reliably execute. Then, a calibration liquid with properties similar to the liquid to be separated is filled into the separation channel, and the separation inlet is controlled to move within the channel at different moving speeds. At the same time, the local flow field disturbance, backflow range, and wall shear changes caused by the movement of the separation inlet are observed through a particle image velocimetry unit. Then, a complete Noctiluca scintillans sample is placed in the channel, and the response of Noctiluca scintillans to the separation zone is recorded at different moving speeds to determine whether the separation zone shifts, stretches, adheres to the wall, aggregates, or the morphology retention coefficient decreases.

[0207] Once the separation suction port reaches the target alignment position, the controller activates the separation suction port to perform the suction operation at a preset suction rate, as shown below: ;in, Indicates the separation of the intake port at time Inhalation rate; The preset aspiration rate is obtained by calibrating the collection efficiency and morphological integrity of intact Noctiluca scintillans entering the separation aspiration port. First, a laminar flow escort flow field is established in the separation device, and the intact Noctiluca scintillans sample is placed within the Noctiluca scintillans response separation zone. Then, the separation aspiration port is controlled to perform aspiration operations at multiple aspiration rates from low to high. The time required for intact Noctiluca scintillans to enter the separation aspiration port, the degree of aspiration path deviation, the co-absorption of impurities, whether the Noctiluca scintillans outline is stretched, whether the morphological retention coefficient decreases, and the integrity rate after collection are recorded respectively.

[0208] During inhalation, the controller acquires the morphology retention coefficient of the complete Noctiluca scintillans from the neural network control model in real time and calculates the instantaneous rate of change of the morphology retention coefficient. If the morphology retention coefficient shows a decreasing trend, that is: The controller then synchronously reduces the suction rate at the separation inlet to mitigate the liquid traction force generated by the suction operation on the intact Noctiluca scintillans.

[0209] The dynamic adjustment of the inhalation rate is expressed as: ;in, This indicates the inhalation rate after dynamic adjustment; This represents the absolute value of the rate at which the shape retention coefficient decreases. Indicates the preset inhalation rate; The morphological change adjustment coefficient is obtained by calibrating the relationship between the rate of decrease of the morphological retention coefficient and the amount of reduction in the inhalation rate. In the calibration experiment, the intact Noctiluca scintillans was placed under different inhalation rates and different laminar flow escort conditions, and the change of its morphological retention coefficient over time was continuously recorded. When a decrease in the morphological retention coefficient was detected, the inhalation rate of the separation inlet was gradually reduced, and it was observed whether the outline of Noctiluca scintillans was restored, whether the area was stable, and whether the response separation zone was restored to its integrity. Then, the amount of reduction in the inhalation rate required to restore Noctiluca scintillans from the morphological decline trend to a stable state under different morphological retention coefficient decrease rates was statistically analyzed to form the adjustment sample data.

[0210] To avoid the inhalation rate being too low, preventing intact *Noctiluca scintillans* from entering the separation inhalation port, the inhalation rate must also meet the following requirements:

[0211] ;

[0212] in, This indicates the inhalation rate after dynamic adjustment; Indicates the preset inhalation rate; The minimum effective intake rate allowed by the separation inlet is calibrated by the minimum flow conditions for Noctiluca scintillation to enter the separation inlet. A stable laminar escort flow field is established in the separation channel, and the intact Noctiluca scintillation is placed within the effective range of the separation inlet. Then, the intake rate of the separation inlet is gradually increased from a lower intake rate. The system continuously records whether the intact Noctiluca scintillation can overcome the effects of local backflow, buoyancy deviation, and axial escort flow to enter the separation inlet. At the same time, the time required for it to enter the inlet, the degree of trajectory deviation, and whether it stagnates or crosses the collection position are recorded.

[0213] When the shape preservation coefficient returns to stability, that is: The controller will either maintain the current inhalation rate or gradually restore it to the preset inhalation rate. The stability threshold of the morphology retention coefficient is determined by statistically analyzing the rate of change of stable and fluctuating morphology samples. Complete *Noctiluca scintillans* samples in normal suspension with intact outlines and stable area changes are selected. The morphology retention coefficient is continuously calculated under different laminar flow rates and inhalation rates, and the rate of change between two adjacent morphology retention coefficients is statistically analyzed to obtain the fluctuation range of the rate of change when the complete *Noctiluca scintillans* is in a stable state. Simultaneously, the rate of change of the morphology retention coefficient is recorded when *Noctiluca scintillans* is subjected to inhalation traction, wall migration, bubble compression, or local flow field disturbances to obtain the range of the rate of change corresponding to the morphological instability state. Subsequently, the boundary value of the rate of change that can distinguish between the stable and instable states is selected as the stability threshold of the morphology retention coefficient.

[0214] The suction operation is no longer a fixed negative pressure suction, but is adjusted in real time according to the morphology of the intact Noctiluca scintillans to maintain the overall structural integrity of the Noctiluca scintillans response separation zone.

[0215] S34, during the inhalation operation, the controller fine-tunes the flow rate of the buffer pump according to the flow rate compensation value of the buffer delivery flow rate, so that a velocity gradient is formed between the axial delivery component and the inhalation rate.

[0216] The flow rate compensation value for buffer escort flow rate, generated by the neural network control model based on the closest distance and relative motion trend between intact Noctiluca scintillans and bubbles and non-luminescent impurities, is... .

[0217] The fine-tuned axial escort component is expressed as follows: ;in, This indicates the axial escort component after fine-tuning; This indicates the axial escort component before fine-tuning; This represents the flow rate compensation value for the buffer escort flow rate.

[0218] The velocity gradient between the axial escort component and the inhalation rate is expressed as: ;in, This indicates the velocity gradient in front of the separation inlet; Indicates the intake rate of the separate intake port; This indicates the effective opening area of ​​the separation inlet; This indicates the axial escort component after fine-tuning; This indicates the distance for calculating the velocity gradient, which is the preset effective distance in front of the separation inlet.

[0219] The effective opening area of ​​the separation inlet is obtained through inlet geometry calibration and actual flow capacity correction. First, the theoretical opening area is determined based on the structural dimensions of the separation inlet. For example, the circular inlet can be calculated based on its inner diameter, while the rectangular or slit-shaped inlet can be calculated based on its width and height. Then, a calibration liquid with viscosity, salinity, and temperature similar to the liquid to be separated is introduced into the separation equipment. The actual flow rate through the inlet is measured under different valve openings and suction negative pressure conditions. The actual flow area involved in the suction is then calculated by combining the average flow velocity in front of the inlet. Due to the presence of boundary layers, local contraction, minor processing deviations, or algal protective filter structures at the edge of the inlet, the actual area that can form a stable suction effect is usually smaller than the theoretical geometric area. Therefore, the theoretical opening area should be compared with the measured flow area, and the flow area that stably sucks up complete Noctiluca scintillans without producing obvious local eddies should be taken as the effective opening area of ​​the separation inlet.

[0220] The velocity gradient calculation distance is determined based on the effective range of the interaction between the suction zone in front of the separation inlet and the buffer escort flow field. During the equipment calibration phase, a laminar escort flow field is first established in the separation channel, and the separation inlet is operated at different suction rates. Subsequently, the location range of streamline curvature, axial velocity change, and lateral repulsion flow field formation in front of the separation inlet is observed using a particle image velocimetry unit. The distance extending upstream from the edge of the inlet to the point where the velocity change significantly weakens is determined, and this distance can be used as a candidate value for the velocity gradient calculation distance. In order to ensure that the velocity gradient can truly reflect the force change of intact Noctiluca scintillans before entering the inlet, it should also be corrected by combining the average equivalent diameter of intact Noctiluca scintillans, the thickness of the response separation zone, and the minimum safe avoidance distance of bubbles or non-luminescent impurities. This ensures that the velocity gradient calculation distance covers the main escort and repulsion area in front of intact Noctiluca scintillans before entering the inlet, rather than being limited to the very near the inlet. The final velocity gradient calculation distance is selected at the location where the velocity gradient change is stable within this range, the interfering substances can be pushed away laterally, and the morphology retention coefficient of intact Noctiluca scintillans does not decrease significantly.

[0221] Under the influence of the velocity gradient, a repulsive flow field biased to the side is formed in front of the separation inlet; this repulsive flow field is expressed as:

[0222] ;

[0223] in, This represents the velocity vector of the repulsive flow field in front of the separation inlet; This represents the conversion factor between the velocity gradient and the repulsive flow field. Represents the velocity gradient; This indicates a unit direction that is biased to the side.

[0224] The conversion coefficient between velocity gradient and repulsive flow field is obtained through local flow field calibration in front of the separation inlet. Before the separation equipment is put into use, a calibration solution with viscosity, salinity and temperature similar to the liquid to be separated is injected into the separation channel. A small amount of tracer particles that can be identified by the particle image velocimetry unit are added to the calibration solution. Then, the buffer pump and the separation inlet are controlled to operate under different axial escort components, different suction rates and different flow rate compensation values. The velocity gradient formed in front of the separation inlet is calculated. At the same time, the particle image velocimetry unit is used to record the lateral streamline deflection velocity, the lateral deflection distance of bubbles or simulated impurity particles and the intensity of the repulsive flow field deflected to the side under the action of the velocity gradient. Then, each set of velocity gradients is paired with the corresponding measured repulsive flow field velocity to form velocity gradient and repulsive flow field calibration data.

[0225] The repulsive flow field is used to push bubbles and non-luminescent impurities away from the intake path of the separation intake port, while intact Noctiluca scintillans enters the separation intake port along the main direction of the velocity gradient under the combined action of the axial escort component and the intake rate.

[0226] The main direction of the intact Noctiluca scintillans entering the separation inlet is indicated as follows: ;in, This indicates the principal direction of the velocity gradient of the intact Noctiluca scintillans as it enters the separation inlet; Indicates the current position of the separation suction inlet; Indicates the representative location of the continuous region containing the complete Noctiluca scintillans; This represents the Euclidean distance.

[0227] To prevent the adjusted escort flow rate from being too strong and disrupting the Noctiluca scintillans response separation zone, the escort flow rate constraint is set as follows:

[0228] ;

[0229] in, This indicates the minimum allowable value for the axial escort component; This indicates the maximum allowable value for the axial escort component; This indicates the axial escort component after fine-tuning.

[0230] The minimum and maximum allowable values ​​of the axial escort component were obtained by calibrating the transport stability and damage risk of intact *Noctiluca scintillans* in the laminar escort flow field. First, a stable *Noctiluca scintillans* response separation zone was formed in the separation channel, and the separation inlet, buffer pump, and collection valve were kept under control. Then, the axial escort component was gradually increased starting from a lower axial escort flow rate. The continuity of the longitudinal movement of intact *Noctiluca scintillans* along the separation channel, the success rate of entering the separation inlet, whether the response separation zone broke, and whether *Noctiluca scintillans* exhibited phenomena such as sedimentation, backflow, crossing the inlet, or being dragged by impurities were recorded. When the axial escort component is too low, intact *Noctiluca scintillans* is difficult to be stably delivered to the separation inlet, or it is easy to stagnate in the local flow field. Therefore, the minimum flow velocity that enables intact *Noctiluca scintillans* to continuously move towards the separation inlet within the predicted time window without significant stagnation is determined as the minimum allowable value of the axial escort component. When the axial escort component continues to increase to a certain level, if situations such as the *Noctiluca scintillans* response separation zone being elongated, the morphology retention coefficient of intact *Noctiluca scintillans* decreasing, wall shear increasing, or bubbles or non-luminescent impurities being carried into the suction path occur, it is considered to be close to the upper limit. The safe flow velocity one level before the occurrence of the above risks is determined as the maximum allowable value of the axial escort component.

[0231] The buffer pump does not simply provide a constant flow rate, but rather creates a local flow field in front of the separation inlet that has both escorting and repelling effects, causing bubbles and non-luminescent impurities to deviate from the suction path, while allowing intact Noctiluca scintillans to smoothly enter the separation inlet.

[0232] S35, during the inhalation operation, the controller controls the collection valve to operate according to the collection valve opening sequence; the collection valve opening sequence is expressed as follows: When the start time of the prediction time window is reached, the controller opens the collection valve, as indicated by:

[0233] ;

[0234] in, Indicates the status of the collection valve; when When, it indicates that the collection valve is open; when When this time, it indicates that the collection valve is closed.

[0235] After the collection valve is opened, the liquid flow containing intact Noctiluca scintillans introduced through the separation suction port is switched to the collection branch, and the resulting initial separated liquid is represented as follows: ;in, Indicates the initial separated liquid; Indicating the collection operation, after the controller determines that the current moment is within the predicted time window corresponding to the opening sequence of the collection valve, it first sends an opening command to the collection valve, causing the downstream passage of the separation suction port to switch from the waste liquid branch or the circulation branch to the collection branch. Subsequently, the separation suction port introduces the complete Noctiluca scintillans in the Noctiluca scintillans response separation zone and a small amount of escort liquid around it into the collection branch, forming a liquid flow containing complete Noctiluca scintillans. After the liquid flow enters the collection branch, it can first enter the buffer collection chamber to reduce the secondary impact of valve switching and pipeline pressure fluctuations on the Noctiluca scintillans, and then flow into the collection container to obtain the initial separation liquid. This indicates that the liquid flow introduced through the separation inlet is activated by the controller after the separation inlet is aligned with the spatial coordinates of the centroid of the complete Noctiluca scintillans distribution, creating a local suction flow field in front of the inlet. At this time, the complete Noctiluca scintillans within the effective range of the inlet, the surrounding buffer solution, and a small amount of liquid to be separated that enter with the flow together constitute the liquid flow introduced through the separation inlet.

[0236] During the opening of the collection valve, the controller continuously calls the photoelectric detection array and particle tracking unit around the separation channel to monitor the attenuation of the luminescence intensity and the amplitude of the floating and sinking speed of Noctiluca scintillans in response to the separation zone in real time.

[0237] The magnitude of luminous intensity attenuation is expressed as: ;in, Indicates time The rate of luminescence intensity decay in the separation band of *Noctiluca scintillans*; This indicates the luminescence intensity of the Noctiluca scintillans response to the separation band at the current moment; This indicates an extremely small positive number that prevents the denominator from being zero; The reference luminescence intensity of the Noctiluca scintillans response separation zone before the collection valve is opened is acquired by a photoelectric detection array during the stable observation phase before the collection valve is opened. When the disturbance is paused, the laminar flow escort flow field is established, the separation inlet is aligned, and the collection valve has not yet been opened, the controller calls the photoelectric detection array outside the separation channel to continuously collect luminescence intensity values ​​at multiple sampling times in the spatial region where the Noctiluca scintillans response separation zone is located. Then, abnormal values ​​caused by instantaneous noise, bubble reflection, local shading, or dark current of the photoelectric detection array are removed, and the remaining luminescence intensity values ​​are averaged or weighted averaged to obtain the reference luminescence intensity of the Noctiluca scintillans response separation zone before the collection valve is opened.

[0238] The amplitude of buoyancy velocity is expressed as: ;in, Indicates time The amplitude of the floating and sinking velocity of *Noctiluca scintillans* in response to the separation zone; This indicates the maximum vertical velocity of the tracked target within the Noctiluca scintillans response separation zone during the current detection period; This represents the minimum vertical velocity of the tracked target within the Noctiluca scintillans response separation zone during the current detection period.

[0239] If the decrease in luminous intensity or the amplitude of floating speed deviates from the preset stable range, the duration of the deviation will begin to accumulate.

[0240] The stability criterion is expressed as follows: ; ;in, and These represent the lower and upper limits of the preset stable range for the magnitude of luminous intensity decay, respectively. and These represent the lower and upper limits of the preset stable range for the amplitude of buoyancy and sinking speed, respectively.

[0241] The lower and upper limits of the preset stable range of luminescence intensity decay were obtained by calibrating the luminescence response changes of intact Noctiluca scintillans during normal collection. First, Noctiluca scintillans samples with high integrity and stable activity were selected. Multiple collection experiments were conducted in the separation device according to the complete process of low-intensity periodic disturbance, laminar flow escort, inlet alignment, and collection valve opening. The baseline luminescence intensity and current luminescence intensity of the Noctiluca scintillans response separation zone before and after the collection valve was opened were recorded, and the luminescence intensity decay range at different collection stages was calculated. Then, the decay range corresponding to the samples that remained intact after collection, were not broken under microscopic observation, and could maintain a reversible luminescence response were taken as the stable sample range. At the same time, the range of abnormal samples with rapid luminescence intensity decrease due to excessive inhalation traction, wall adhesion aggregation, bubble interference, or damage to Noctiluca scintillans was recorded.

[0242] The lower and upper limits of the preset stable range of floating and sinking velocity amplitudes were obtained by calibrating the motion stability of the Noctiluca scintillans response separation zone during laminar flow escort and aspiration collection. First, multiple rounds of testing were conducted on intact Noctiluca scintillans samples under formal separation conditions. The vertical velocity change of the tracked target in the Noctiluca scintillans response separation zone was continuously recorded using a particle tracking unit, and the floating and sinking velocity amplitudes were calculated in each detection cycle. For samples with high integrity after collection, no wall migration, no breakage of the response separation zone, and successful entry into the collection branch, their floating and sinking velocity amplitudes were used as stable sample data. For abnormal processes caused by rapid bubble rise, impurity dragging, excessive suction at the aspiration port, Noctiluca scintillans settling or wall aggregation, their floating and sinking velocity amplitudes were used as unstable sample data.

[0243] When the duration of deviation from the preset stable range exceeds the allowable deviation time, the Noctiluca scintillans response separation band is considered unstable, as indicated by: ;in, This indicates the duration during which the luminous intensity decay or the floating / sinking speed deviates from the preset stable range. Indicates the allowable deviation duration.

[0244] The duration for which the luminescence intensity decay or the floating-sinking velocity amplitude deviates from the preset stable range is obtained by the controller through time accumulation of continuous monitoring results. During the period when the collection valve is open, the controller continuously reads the luminescence intensity decay and floating-sinking velocity amplitude of the Noctiluca scintillans response separation zone according to the preset sampling cycle, and determines whether they fall into the corresponding preset stable range. When any indicator exceeds the preset stable range for the first time, the moment is recorded as the deviation start time, and the deviation time is accumulated.

[0245] The allowable deviation duration was obtained through calibration experiments. Specifically, the duration difference between short-term fluctuations and actual instability was recorded during normal and abnormal collection processes. The maximum recoverable deviation duration caused by sampling noise, short-term flow velocity fluctuations, or brief occlusion during normal collection was used as the basic reference. At the same time, it was corrected by combining the time characteristics of continuous decay of luminescence intensity, continuous abnormal floating and sinking velocity, or the time before the response separation zone broke in the actual unstable samples. Finally, the time value that can filter short-term recoverable fluctuations and trigger protective actions before the Noctiluca scintillans is significantly damaged or adheres to the wall was selected as the allowable deviation duration.

[0246] When the separation zone of *Noctiluca scintillans* is determined to be unstable, the controller immediately triggers a protective action, suspending the suction operation at the separation inlet and closing the collection valve. ;in, Indicates the intake rate of the separate intake port; This indicates the status of the collection valve.

[0247] The controller reapplies low-intensity periodic perturbation to the liquid to be separated, reacquires the reversible luminescence response and slow buoyancy response, and reconstructs the Noctiluca scintillans response separation zone; after the Noctiluca scintillans response separation zone is reconstructed, a new response-maintaining separation control command is generated, and collection continues; if the instability protection is not triggered, the controller continues to execute the aspiration and collection process until the preset total collection volume or preset collection time is met.

[0248] The collection termination condition is expressed as follows: ;or: ;in, Indicates the current cumulative collected volume; Indicates the preset total collection amount; Indicates the current cumulative collection time; This indicates the preset collection time.

[0249] The preset total collection volume is determined based on experimental requirements, the number of target Noctiluca scintillans, the throughput of the separation equipment, and the time that Noctiluca scintillans remains alive. If the subsequent use is for microscopic observation, activity detection, or small-volume culture, the volume of liquid to be collected is estimated based on the required number of Noctiluca scintillans individuals and the concentration of Noctiluca scintillans per unit volume, and this volume is set as the preset total collection volume. If the subsequent use is for amplification culture or multiple sets of experimental retests, the preset total collection volume can be appropriately increased, but it should still be ensured that the activity of the Noctiluca scintillans separation solution will not decrease due to prolonged residence, local hypoxia, or shear accumulation during the collection process.

[0250] The preset collection time is determined based on the effective inhalation rate of the separation inlet, the flow rate of the collection branch, and the time during which the Noctiluca scintillans response separation zone can be stably maintained. That is, under the premise of ensuring that the Noctiluca scintillans response separation zone does not significantly decay, does not adhere to the wall and aggregates, and meets the integrity requirements, the shortest or moderate time that can achieve the target collection amount is selected.

[0251] When any collection termination condition is met, the controller closes the collection valve and represents the liquid in the collection branch as the final Noctiluca scintillans separation liquid as follows: ;in, This indicates the final Noctiluca scintillans separation solution; This indicates the completion of the operation of collecting and outputting the final separated liquid; This refers to the initial separated liquid obtained during the collection process.

[0252] The process of collecting and outputting the final separated liquid follows this sequence: termination determination, valve and pump shutdown, branch purging, buffer storage, status confirmation, and output sealing. When the controller determines that the cumulative collection volume has reached the preset total collection volume, the cumulative collection time has reached the preset collection time, or the Noctiluca scintillans response to the separation zone instability protection trigger condition, it first stops the suction operation at the separation inlet and switches the collection valve from the open to the closed state to prevent subsequent separated liquid, air bubbles, or non-luminescent impurities from continuing to enter the collection branch. Then, it controls the buffer pump to flush the inlet section of the collection branch at a low flow rate for a short time, ensuring that the liquid that has entered the collection branch but has not yet reached the collection container is completely purified. Throughout the night, the Noctiluca scintillans liquid continues to be gently pushed into the collection container, while avoiding the generation of new strong shear disturbances. Then, the buffer pump is turned off or switched to a microfluidic state to maintain the liquid level, so that the initial separated liquid in the collection container is in a static buffer state. The photoelectric detection array or the particle detection unit at the inlet of the collection container confirms that there are no more obvious intact Noctiluca scintillans passing through the end liquid flow. After the effective liquid flow in the collection branch has completely entered the collection container, the controller reads the liquid level, volume or weight information of the collection container to confirm that it meets the preset collection requirements. The liquid in the collection container is then mixed with low disturbance for a short time or allowed to stand for equilibrium, so that the intact Noctiluca scintillans is evenly dispersed in the collection liquid.

[0253] In summary, separation is not performed according to a fixed pump and valve sequence. Instead, under the constraints of generated response-based separation control commands, it sequentially completes disturbance suspension, laminar flow escort, anti-wall adhesion control, inlet alignment, dynamic intake, interference rejection, collection valve opening and closing, and instability protection. When the Noctiluca scintillans response separation zone becomes unstable, the equipment can automatically stop intake and re-execute the above steps, thereby obtaining the final Noctiluca scintillans separation solution while ensuring the integrity of the live Noctiluca scintillans and the stability of the response separation zone structure.

[0254] Example 2

[0255] like Figure 3 As shown, a control system for a neural network-based separation device for Noctiluca scintillans includes the following modules:

[0256] A disturbance response construction module is used to control the disturbance unit to apply low-intensity periodic disturbances to the liquid to be separated containing Noctiluca scintillans, so that Noctiluca scintillans produces a reversible luminescence response and a slow floating and sinking response in the separation channel.

[0257] The response data acquisition module is used to acquire multiple frames of two-dimensional spatial distribution images or light intensity-displacement coupling data of the Noctiluca scintillans response separation zone within a continuous time window, and form a three-dimensional spatiotemporal feature tensor according to the acquisition sequence.

[0258] The neural network control module is used to input the three-dimensional spatiotemporal feature tensor into the neural network control model, distinguish between intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles and non-luminescent impurities, output the morphology retention coefficient of intact Noctiluca scintillans, and generate response-maintaining separation control commands.

[0259] The laminar flow escort execution module is used to control the disturbance unit to stop outputting periodic driving force according to the disturbance pause time in the response-maintaining separation control command, and to control the buffer pump to form a laminar flow escort flow field and a radial anti-wall flow field.

[0260] The inhalation port adjustment module is used to control the separation inhalation port to align with the center of gravity of the complete Noctiluca scintillans distribution according to the separation inhalation port alignment position in the response-maintaining separation control command, and to dynamically adjust the inhalation rate according to the instantaneous change rate of the morphology retention coefficient.

[0261] The collection valve control module is used to control the collection valve to open or close according to the collection valve opening sequence in the response-maintaining separation control command, so that the liquid flow containing intact Noctiluca scintillans introduced through the separation inlet enters the collection branch.

[0262] An instability protection module is used to monitor the luminescence intensity decay and buoyancy velocity amplitude of the Noctiluca scintillans response separation zone during the collection process. When the duration of the luminescence intensity decay or buoyancy velocity amplitude deviating from the preset stable range exceeds the allowable deviation time, the Noctiluca scintillans response separation zone is determined to be unstable, and the separation inlet is paused, the collection valve is closed, and the Noctiluca scintillans response separation zone is rebuilt, until the preset total collection volume or preset collection time is met before the Noctiluca scintillans separation liquid is output.

[0263] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0264] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A control method for a neural network-based separation device for Noctiluca scintillans, characterized in that, Includes the following steps: The separation device applies low-intensity periodic disturbances to the liquid to be separated containing Noctiluca scintillans, causing the Noctiluca scintillans to produce a reversible luminescence response and a slow floating and sinking response within the separation channel. Based on the reversible luminescence response and slow buoyancy response, a separation zone for Noctiluca scintillans response is formed; Based on the Noctiluca scintillans response separation zone input neural network control model, the neural network control model distinguishes between intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles, and non-luminescent impurities; Obtain a response-holding separation control command, which is used to determine the disturbance pause time, the separation inlet alignment position, the buffer escort flow rate, and the collection valve opening sequence; Based on the aforementioned response-maintaining separation control command, the disturbance unit, separation inlet, buffer pump, and collection valve are controlled to operate in a coordinated manner. When the Noctiluca scintillans response separation zone remains intact and does not adhere to the wall, the intact Noctiluca scintillans is introduced into the collection branch to obtain the Noctiluca scintillans separation solution.

2. The control method for a neural network-based separation device for Noctiluca scintillans according to claim 1, characterized in that, The perturbation unit of the separation device outputs a periodic driving force in the form of a sine wave. The frequency of the periodic driving force is 0.5Hz to 5Hz, and the fluid shear force corresponding to its amplitude does not exceed the elastic deformation threshold of the Noctiluca scintillans cell membrane. ;in, Indicates at time Fluid shear force formed by the periodic driving force acting on the liquid to be separated; This represents the elastic deformation threshold of the cell membrane of Noctiluca scintillans, which is used to apply the low-intensity periodic perturbation to the liquid to be separated, driving Noctiluca scintillans to make restricted floating and sinking movements in the vertical direction within the separation channel. Throughout the application of the low-intensity periodic perturbation, a photoelectric detection array arranged around the separation channel is activated to continuously collect photon counts or luminescence intensity values ​​in each region of the liquid to be separated, and obtain the time-series curve of the luminescence intensity of each region changing with time, which serves as the reversible luminescence response.

3. The control method for a neural network-based separation device for Noctiluca scintillans according to claim 2, characterized in that, The particle image velocimetry unit, arranged on the side of the separation channel, is activated synchronously to record the displacement of particles in the liquid to be separated in the vertical direction in real time, which is expressed as: ;in, Indicates the first The tracked target was in the first Vertical displacement within each sampling interval; Indicates the first The tracked target was in the first Vertical position at each sampling time; Indicates the first The tracked target was in the first The vertical position at each sampling moment is determined, and its instantaneous buoyancy and sinking speed and speed change period are calculated as the slow buoyancy response. The time-series curve of the reversible luminescence response is spatiotemporally coupled and matched with the velocity change period of the slow buoyancy response to extract a continuous liquid layer region that simultaneously satisfies the following conditions: the peak luminescence intensity and the perturbation phase difference are constant, and the buoyancy velocity amplitude is within a preset range characteristic of living algae. The spatial position of this continuous liquid layer region in the separation channel is dynamically calibrated as the Noctiluca scintillans response separation zone, represented as follows: ;in, Indicates time The separation band of Noctiluca scintillation response obtained by dynamic calibration; This indicates a spatial connectivity filtering operation; Indicates the first Each detection area belongs to a candidate response area; Indicates the first The spatial location of each detection area in the separation channel.

4. The control method for a neural network-based separation device for Noctiluca scintillans according to claim 1, characterized in that, Acquire multiple frames of two-dimensional spatial distribution images or light intensity-displacement coupled data of the Noctiluca scintillans response separation band within a continuous time window, and stack the multiple frames of data into a three-dimensional spatiotemporal feature tensor according to the acquisition time sequence. Before stacking, normalize each input frame as follows: ;in, This represents the normalized input value; This represents the input value before normalization; Indicates the first The average of the data channels; Indicates the first Standard deviation of each data channel; This indicates an extremely small positive number that prevents the denominator from being zero; Indicates the position along the length direction in the coordinate system of the separation channel; Indicates the position in the height direction within the separation channel coordinate system; This indicates the data channel number, and the three-dimensional spatiotemporal feature tensor is used as the input data for the neural network control model; The neural network control model sequentially performs spatiotemporal feature extraction and encoding-decoding semantic segmentation on the three-dimensional spatiotemporal feature tensor, and outputs classification confidence values ​​pixel-by-pixel or connected-component-by-connection as follows: ;in, Indicates spatial location Belongs to the Classification confidence of the target class; Indicates spatial location Belongs to the Classification score of the target category; This represents the total number of target categories. Based on the classification confidence level, each target is classified into intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles, and non-luminescent impurities. Simultaneously, the neural network control model outputs a morphology retention coefficient for the continuous region containing intact Noctiluca scintillans, characterizing its outline integrity and area stability, expressed as: ;in, Indicates time Morphological preservation coefficient of the continuous region containing the complete Noctiluca scintillans; Indicates the completeness of the outline; Indicates area stability; The weighting coefficients are represented by the area stability of the continuous region containing the intact Noctiluca scintillans. ;in, Indicates time Area stability of the continuous region containing the complete Noctiluca scintillans; This represents the area of ​​the continuous region containing the complete Noctiluca scintillans at the current moment; It represents the average area of ​​the continuous region containing the complete Noctiluca scintillans within a preset time range prior to the current moment.

5. The control method for a neural network-based separation device for Noctiluca scintillans according to claim 4, characterized in that, The neural network control model, based on the distribution centroid and density maxima location of intact *Noctiluca scintillans*, represents the distribution centroid of the continuous region containing intact *Noctiluca scintillans* as follows: ;in, Indicates time The distribution center of the continuous area where the complete Noctiluca scintillans is located; Indicates the continuous region containing the complete Noctiluca scintillans; Indicates spatial location At any moment The classification confidence level of *Noctiluca scintillans* belonging to the complete *Noctiluca scintillans*; Indicates spatial location Position vector in the separate channel coordinate system, spatial position The position vector in the separation channel coordinate system is used to generate the alignment position of the separation inlet; the rate of change of the shape retention coefficient over time is expressed as follows: ;in, Indicates time The rate of change of the lower morphology retention coefficient over time; This represents the shape preservation coefficient at the current moment; This represents the shape preservation coefficient at the previous shape detection time. The time interval between two consecutive calculations of the morphology retention coefficient is represented, and the disturbance pause time is generated before the morphology retention coefficient drops to a preset threshold. Based on the closest distance and relative movement trend between the intact *Noctiluca scintillans* and the bubbles and non-luminescent impurities, a flow rate compensation value for the buffer escort flow rate is generated to push away interfering materials during the escort process. Based on the current longitudinal displacement velocity and distance from the collection port of the intact *Noctiluca scintillans*, a predicted time window for its arrival at the collection port is calculated, and the collection valve opening sequence is generated based on the predicted time window as follows: ;in, Indicates the opening sequence of the collection valve; Indicates the moment the collection valve opens; Indicates the time when the collection valve closes.

6. The control method for a neural network-based separation device for Noctiluca scintillans according to claim 1, characterized in that, The response-maintaining separation control command is analyzed to extract the disturbance pause time, the separation inlet alignment position, the buffer escort flow rate setpoint, and the collection valve opening sequence. The flow rate setpoint is then decomposed into an axial escort component along the separation channel axis and a radial anti-wall-adhesion component along the separation channel radially, as follows: ;in, This indicates the flow rate setpoint for buffer escort flow; This indicates the axial escort component along the axis of the separation channel; This represents the radial anti-wall component along the radial direction of the separation channel; The control of the disturbance unit to stop outputting periodic driving force according to the disturbance pause time is expressed as follows: ;in, Indicates the time of the disturbance element. The periodic driving force output; This indicates the moment when the disturbance is paused, allowing the flow field within the separation channel to return to a laminar escort state. Simultaneously, the buffer pump is started to establish a laminar escort flow field along the longitudinal direction of the separation channel using the axial escort component. At the same time, driven by the radial anti-wall component, a micro-pressure difference or micro-eddy current is generated near the wall of the separation channel towards the center of the channel, thereby inhibiting the Noctiluca scintillans from migrating and accumulating towards the wall in response to the separation zone.

7. The control method for a neural network-based separation device for Noctiluca scintillans according to claim 6, characterized in that, After the laminar flow escort flow field is stabilized, the separation inlet is moved to the spatial coordinates of the centroid of the intact Noctiluca scintillans distribution, according to the alignment position of the separation inlet, and the separation inlet is activated to perform the inhalation operation at a preset inhalation rate. During the inhalation process, the morphology retention coefficient of the intact Noctiluca scintillans output by the neural network control model is acquired in real time, and the preset inhalation rate is dynamically adjusted according to the instantaneous rate of change of the morphology retention coefficient, expressed as: ； in, This indicates the inhalation rate after dynamic adjustment; This represents the absolute value of the rate at which the shape retention coefficient decreases. Indicates the preset inhalation rate; The morphology change adjustment coefficient is indicated. If the morphology retention coefficient shows a decreasing trend, the inhalation rate is reduced simultaneously to slow down the liquid flow traction until the morphology retention coefficient returns to stability, so as to maintain the overall structural integrity of the Noctiluca scintillans response separation zone.

8. The control method for a neural network-based separation device for Noctiluca scintillans according to claim 7, characterized in that, During the inhalation operation, the buffer pump is fine-tuned according to the flow rate compensation value of the buffer escort flow rate, so that a velocity gradient is formed between the axial escort component and the inhalation rate, expressed as: ;in, This indicates the velocity gradient in front of the separation inlet; Indicates the intake rate of the separate intake port; This indicates the effective opening area of ​​the separation inlet; This indicates the axial escort component after fine-tuning; The velocity gradient is used to calculate the distance, generating a repulsive flow field biased to the side in front of the separation inlet. This repulsive flow field pushes the bubbles and non-luminescent impurities away from the intake path of the separation inlet, while simultaneously guiding the intact Noctiluca scintillans smoothly into the separation inlet along the main direction of the velocity gradient. The main direction of the intact Noctiluca scintillans entering the separation inlet is represented as: ;in, This indicates the principal direction of the velocity gradient of the intact Noctiluca scintillans as it enters the separation inlet; Indicates the current position of the separation suction inlet; Indicates the representative location of the continuous region containing the complete Noctiluca scintillans; This represents the Euclidean distance.

9. The control method for a neural network-based separation device for Noctiluca scintillans according to claim 8, characterized in that, The collection valve is controlled to open at the beginning of the predicted time window according to the opening sequence, so that the liquid flow containing the intact Noctiluca scintillans introduced through the separation inlet is switched to the collection branch to obtain the initial separation liquid; during the opening of the collection valve, the photoelectric detection array and particle tracking unit around the separation channel are continuously activated to monitor the luminescence intensity attenuation and floating / sinking velocity amplitude of the Noctiluca scintillans in response to the separation zone in real time; the luminescence intensity attenuation amplitude is expressed as: ;in, Indicates time The rate of luminescence intensity decay in the separation band of *Noctiluca scintillans*; This indicates the luminescence intensity of the Noctiluca scintillans response to the separation band at the current moment; This indicates an extremely small positive number that prevents the denominator from being zero; The reference luminescence intensity of the luminescent algae response separation zone is represented by: (The luminescence intensity is represented by:) ;in, Indicates time The amplitude of the floating and sinking velocity of *Noctiluca scintillans* in response to the separation zone; This indicates the maximum vertical velocity of the tracked target within the Noctiluca scintillans response separation zone during the current detection period; This represents the minimum vertical velocity of the tracked target within the Noctiluca scintillans response separation zone during the current detection period; If the duration of the decrease in luminescence intensity or the amplitude of the floating and sinking velocity deviating from the preset stable range exceeds the allowable deviation time, it is determined that the Noctiluca scintillans response separation zone is unstable, and a protection action is immediately triggered—the suction operation of the separation inlet is suspended and the collection valve is closed, the Noctiluca scintillans response separation zone is re-executed, and collection is resumed after the reconstruction is completed until the preset total collection amount or collection time is met, at which point the collection valve is closed to obtain the final Noctiluca scintillans separation liquid.

10. A control system for a neural network-based separation device for *Noctiluca scintillans*, used to implement the control method for the neural network-based separation device for *Noctiluca scintillans* as described in any one of claims 1-9, characterized in that, Includes the following modules: The disturbance response construction module is used to control the disturbance unit to apply low-intensity periodic disturbance to the liquid to be separated containing Noctiluca scintillans, so that Noctiluca scintillans generates a reversible luminescence response and a slow floating and sinking response in the separation channel, and forms a Noctiluca scintillans response separation zone based on the reversible luminescence response and the slow floating and sinking response. The response data acquisition module is used to acquire multiple frames of two-dimensional spatial distribution images or light intensity-displacement coupling data of the Noctiluca scintillans response separation zone within a continuous time window, and form a three-dimensional spatiotemporal feature tensor according to the acquisition sequence. The neural network control module is used to input the three-dimensional spatiotemporal feature tensor into the neural network control model, distinguish between intact Noctiluca scintillans, damaged Noctiluca scintillans, bubbles and non-luminescent impurities, output the morphology retention coefficient of intact Noctiluca scintillans, and generate response-maintaining separation control commands. The laminar flow escort execution module is used to control the disturbance unit to stop outputting periodic driving force according to the disturbance pause time in the response-maintaining separation control command, and to control the buffer pump to form a laminar flow escort flow field and a radial anti-wall flow field. The inhalation port adjustment module is used to control the separation inhalation port to align with the center of gravity of the complete Noctiluca scintillans distribution according to the separation inhalation port alignment position in the response-maintaining separation control command, and to dynamically adjust the inhalation rate according to the instantaneous change rate of the morphology retention coefficient. The collection valve control module is used to control the collection valve to open or close according to the collection valve opening sequence in the response-maintaining separation control command, so that the liquid flow containing intact Noctiluca scintillans introduced through the separation inlet enters the collection branch. An instability protection module is used to monitor the luminescence intensity decay and buoyancy velocity amplitude of the Noctiluca scintillans response separation zone during the collection process. When the duration of the luminescence intensity decay or buoyancy velocity amplitude deviating from the preset stable range exceeds the allowable deviation time, the Noctiluca scintillans response separation zone is determined to be unstable, and the separation inlet is paused, the collection valve is closed, and the Noctiluca scintillans response separation zone is rebuilt, until the preset total collection volume or preset collection time is met before the Noctiluca scintillans separation liquid is output.

Images