Intelligent bionic seduction and investigation robot system and working method thereof

Abstract

The application discloses an intelligent bionic estrus induction and estrus detection robot system and a working method thereof, and the system comprises: an autonomous mobile inspection platform serving as a physical carrier of the estrus induction and estrus detection robot; an anti-habituation bionic excitation subsystem for generating a modulation control sequence and dynamically modulating a pre-recorded pig audio sample to output an anti-habituation bionic estrus induction audio, and combining a smell pulse to trigger a bionic estrus induction mechanism; a multi-modal perception and feature extraction subsystem for perceiving physiological data and behavioral reaction data of sows after being induced by the bionic estrus induction mechanism, and converting the data into multi-dimensional response features; and a large model diagnosis subsystem based on cross attention, for cross attention fusion of the multi-dimensional response features, obtaining an estrus diagnosis result according to the fused features, and driving the system to perform a closed-loop execution stage. The application effectively replaces a test estrus boar, and significantly reduces non-production days of breeding.

Description

Technical Field

[0001] This invention belongs to the field of novel aquaculture technology, specifically relating to an intelligent biomimetic estrus-inducing and estrus-detecting robot system and its working method. Background Technology

[0002] Against the backdrop of global agricultural modernization, the pig farming industry is undergoing a profound transformation from extensive to intensive, digitalized, and intelligent operations. Within this complex biological asset production system, the management efficiency of breeding sow herds forms the cornerstone of the entire industry chain's economic benefits. The management of the sow reproductive cycle, particularly the accuracy and timeliness of estrus detection, directly determines the number of non-productive days, conception rate, and the number of weaned piglets produced annually. Authoritative industry data shows that each additional non-productive day not only means ineffective feed cost input (approximately 2.5-3.0 kg of feed consumed daily), but also involves increased fixed asset depreciation and potential opportunity cost losses. In ultra-large-scale pig farms with tens of thousands of pigs, a 1% decrease in estrus detection rate can result in direct economic losses of millions of RMB. Therefore, accurate and efficient estrus detection technology is crucial for improving farming efficiency.

[0003] From a microscopic perspective of biophysiology, estrus in sows is a complex physiological process finely regulated by the hypothalamic-pituitary-ovarian axis. A typical estrous cycle averages 21 days (range 18-24 days), with the estrous phase lasting only about 40 to 60 hours. The optimal insemination window is even shorter, typically occurring within 12 to 24 hours after the onset of the standing reflex. During this critical period, estrogen levels in the sow rise sharply, triggering a series of significant physiological and behavioral changes. These changes include: noticeable redness and swelling of the vulva and mucus discharge; increased seeking behavior towards boars, manifested as frequent, specific courtship calls; and most importantly, the standing reflex—the characteristic posture of a sow remaining motionless with stiff limbs, erect ears, and a boar's presence when her back is subjected to pressure or she senses it. This reflex is the "gold standard" for determining whether a sow has entered her breeding season.

[0004] However, with the continuous advancement of genetic breeding technology, while modern high-yielding sow breeds pursue economic traits such as high litter size, the intensity of their estrus behavior is showing a worrying trend of "silent estrus." Studies have shown that in modern intensive high-density farming environments, due to factors such as limited space in gestation crates, insufficient lighting, high ammonia concentrations, and lack of boar stimulation, approximately 15% to 30% of sows exhibit the so-called "silent estrus" phenomenon. Although these sows ovulate normally and are physiologically capable of conception, their external behavioral signs are extremely weak, or even completely lacking the traditional standing reflex or redness and swelling characteristics. This phenomenon exposes traditional detection methods relying on visual observation and experience to an unprecedented risk of failure, easily leading to missed matings and wasting valuable reproductive cycles.

[0005] To address these challenges, the industry has developed various detection methods, primarily covering traditional manual investigation methods, IoT monitoring technology based on contact sensors, non-contact monitoring technology based on computer vision, and the emerging auxiliary investigation robot technology. However, a deeper analysis of these existing technologies reveals numerous insurmountable pain points, technical bottlenecks, and legal risks in practical applications.

[0006] Currently, the vast majority of pig farms still use the traditional "manual intervention + estrus-testing boar" operating model. This model requires experienced technicians to guide the estrus-testing boar to patrol in front of the sow pens at least twice a day, and to manually press the back of each sow to induce the standing reflex. This method has obvious drawbacks:

[0007] First, there is the irreconcilable contradiction between labor intensity and efficiency. In large pig farms with tens of thousands of pigs, manually checking each pig for symptom is an enormous physical task, easily leading to physiological fatigue among staff. As working hours increase, the sharpness of human observation and the standard of pressure applied inevitably decline, causing the missed detection rate to rise significantly in the later stages of the operation. In addition, this high-intensity repetitive labor also leads to a high turnover rate of farm workers, further exacerbating the labor shortage.

[0008] Secondly, there is a significant reliance on subjective judgment. The accuracy of estrus detection depends heavily on the technician's personal experience and sense of responsibility. For sows where early signs of estrus are not obvious, newly hired or inexperienced employees often fail to detect the fleeting ear-raising movement or subtle changes in body temperature, thus missing the optimal mating opportunity. This "empirical" judgment method is difficult to quantify and standardize, hindering data-driven management in pig farms.

[0009] Thirdly, there are biosecurity risks. As a live biological vector, the test boar frequently moves between different buildings and pens, and its oral-nasal contact and droplet transmission pose a high-risk route for cross-infection of highly contagious diseases such as porcine reproductive and respiratory syndrome (PRRS) and African swine fever. Furthermore, the raising, herding, and management of boars are not only costly, but the boars themselves are also aggressive, posing a potential threat to the personal safety of on-site workers.

[0010] To achieve automated monitoring, some existing technologies use ear tags, collars, or implanted sensors to collect data on sows' body temperature, activity levels, or vaginal resistance.

[0011] First, stress is the primary challenge facing this type of technology. Pigs, as sensitive animals, have a natural aversion to foreign objects worn on their bodies. The fright during the restraint and installation process, as well as the long-term physical pressure from the equipment, can easily trigger stress in sows, leading to elevated cortisol levels. This, in turn, interferes with normal reproductive hormone secretion and inhibits normal estrus expression. For example, ear tag installation may cause ear infections, and collars may cause skin abrasion.

[0012] Secondly, environmental adaptability and maintenance costs are significant issues. The internal environment of pigsties is extremely harsh, with high humidity, high concentrations of ammonia and dust, as well as the scratching and biting behavior of pigs, resulting in a high failure rate of precision electronic equipment. In large-scale farming scenarios, the labor and material costs associated with frequent battery replacements and repairs of detached equipment make the commercial promotion of this technology solution extremely difficult.

[0013] Finally, single-modal data has a high false alarm rate. For example, relying solely on accelerometers to monitor activity levels makes it difficult to distinguish between agitation caused by estrus and increased activity due to hunger, fighting in groups, or environmental stress, leading to a large number of false positive alarms. Although some studies have attempted to combine body temperature data, body temperature is greatly affected by factors such as ambient temperature, food intake, and disease, making single-threshold determination still inaccurate.

[0014] With the development of artificial intelligence, non-contact monitoring based on computer vision has become a hot topic, and simple "boar carts" or bionic robots have appeared on the market. However, these technologies have serious flaws in their underlying logic.

[0015] First, the lack of active stimulation leads to missed detections. Existing visual systems are mostly "passive observers," meaning they only record and analyze the sow's natural behavior through cameras. However, biological characteristics dictate that the sow's standing reflex is a typical "stimulus-response" behavior. In the absence of a boar providing multi-sensory stimulation—visual, olfactory (sex pheromones), auditory (courtship calls), and tactile—sows in silent estrus or early estrus often do not exhibit obvious standing characteristics. Passive visual monitoring systems are almost helpless against such sows, resulting in a very high rate of missed detections.

[0016] Second, there is the challenge of the biological filtering mechanism of the reticular activating system (RAS). Existing biomimetic devices mostly employ linear pseudo-random number generators (PRNGs) or simple frequency scanning strategies. Neurobiological research indicates that the reticular activating system in the mammalian brainstem possesses a highly efficient "periodic filtering mechanism." Because the signals generated by current technologies still exhibit periodic or simple linear patterns in their mathematical nature, they are easily recognized as invalid background noise by the auditory center of sows, leading to an exponential decline in long-term induction rates, i.e., severe neural habituation.

[0017] Third, there is the challenge of constructing a diffusion field based on Fick's law. Existing spray technologies ignore the physical law that gas molecule concentration decreases non-linearly with increasing propagation distance. This "open-loop flooding" release leads to a severe "distance-concentration" mismatch: when sows approach the nozzle, the localized high concentration causes rapid saturation of olfactory receptor proteins and triggers desensitization; when sows move away, the concentration falls below the arousal threshold. It is impossible to establish a constant and effective stimulation field in the dynamically changing location of the pigs.

[0018] Fourth, insufficient depth of multimodal fusion and "domain shift". Existing research is mostly limited to a single modality (visual or auditory only) or uses simple decision-level fusion (such as weighted voting). This shallow fusion approach cannot capture the deep semantic relationships between modalities. At the same time, existing visual models are highly sensitive to environmental background. Different lighting conditions in pigsties, fence materials, ground stains, and differences in pig size and skin color can all cause the model's performance to drop sharply when applied across different scenarios. Especially at night or in low-light conditions, a single vision system often fails completely.

[0019] In summary, existing estrus detection technologies for pigs have significant technological gaps in areas such as resistance to behavioral changes caused by active stimulation, precise control of odor release, deep fusion of multimodal data, and robustness in complex environments. There is an urgent need to develop a robotic system capable of simulating multidimensional random variations in stimuli from real boars to overcome biological adaptation, and capable of deeply integrating visual, thermal infrared, and acoustic information for comprehensive intelligent judgment. Summary of the Invention

[0020] To address the aforementioned shortcomings of existing technologies, the intelligent biomimetic estrus-inducing robot system and its working method provided by this invention solve the problems of high labor costs, significant biosecurity risks, and missed detection of silent estrus that arise from the reliance on manual traction of boars for estrus testing in existing pig farms.

[0021] To achieve the aforementioned objectives, the technical solution adopted by this invention is: an intelligent biomimetic arousal-inducing and arousal-detecting robot system, comprising:

[0022] The autonomous mobile inspection platform serves as the physical carrier of the seduction and seduction robot;

[0023] The anti-habit biomimetic excitation subsystem is used to generate modulation control sequences and dynamically modulate pre-recorded pig audio samples to output anti-habit biomimetic estrus-inducing audio, and combine it with odor pulses to trigger a biomimetic estrus-inducing mechanism.

[0024] The multimodal perception and feature extraction subsystem is used to perceive the physiological and behavioral response data of sows after being stimulated by the biomimetic estrus mechanism, and convert them into multidimensional response features.

[0025] The large model diagnostic subsystem based on cross-attention is used to perform cross-attention fusion on multi-dimensional response features, obtain estrus diagnosis results based on the fused features, and drive the system to perform closed-loop execution.

[0026] Furthermore, the autonomous mobile inspection platform includes a functional structure, a mobile-adaptive chassis, a navigation module, and an identity binding module;

[0027] The functionalized external structure adopts a static biomimetic boar shell and is equipped with an acoustic waveguide, a hidden odor nozzle, and a ranging sensor.

[0028] The mobile adapter chassis, equipped with a functional structure, moves within the pigsty.

[0029] The navigation module uses laser SLAM to build a map and combines it with UWB tags to navigate the movement and stopping of the estrus detection robot in the pigsty;

[0030] After the estrus detection robot stops, the identity binding module collects multimodal response data by reading the sow's ear tag with RFID and binds it to the individual's identity.

[0031] Furthermore, the anti-habit bionic excitation subsystem includes an anti-habit hyperchaotic sound field generation module and a closed-loop control loop based on physical law constraints, deployed in the edge computing unit of the erotic detection robot.

[0032] The anti-habit hyperchaotic sound field generation module is equipped with a hyperchaotic dynamics model. It generates a modulation control sequence through a four-dimensional hyperchaotic system and applies the modulation control sequence to pre-recorded pig audio samples to output anti-habit biomimetic estrus-inducing audio with high biosimilarity and unpredictability in the time and frequency domains.

[0033] The closed-loop control circuit based on physical constraints uses the real-time distance between the odor nozzle of the estrus detection robot and the sow's snout as the input of the inverse model. It calculates the odor stimulation intensity required to reach the effective stimulation threshold of the sow's snout and dynamically outputs a PWM control signal with a corresponding duty cycle based on the change in real-time distance to drive the solenoid valve of the odor nozzle that releases hormones, thereby maintaining the effective odor stimulation intensity of the target area in the pig house environment.

[0034] Furthermore, in the anti-habitual hyperchaotic sound field generation module, the dynamic equations for generating the control signal are:

[0035]

[0036] In the formula, For the state variables of a four-dimensional hyperchaotic system, Representing state variables respectively Rate of change over time; , , , These are system control parameters or coupling parameters used to adjust the coupling strength and nonlinear evolution characteristics between various state variables of the system. The bias term is used to drive the continuous evolution of the fourth-dimensional state variable; wherein, the state variable is mapped to form an audio modulation control quantity, which is used to dynamically modulate one or more of the fundamental frequency parameters, amplitude envelope parameters, formant parameters and rhythm parameters of the pre-recorded pig audio samples.

[0037] Furthermore, during pre-training, the inverse model uses the compensation relationship between odor stimulus intensity and distance as a training constraint to improve the stability and rationality of the inverse model's output.

[0038] Furthermore, the multimodal perception and feature extraction subsystem includes a multimodal response data synchronous acquisition module and a multimodal feature quantization module;

[0039] The multimodal response data collected by the multimodal response data synchronous acquisition module includes visible light image sequences, surface infrared temperature data, audio response data, and three-dimensional depth data for sows.

[0040] The multimodal feature quantization module constructs multidimensional response features, including visual features, thermal infrared features, and acoustic features, after complex background removal and mask extraction of the collected multimodal response data.

[0041] Furthermore, in the multimodal feature quantization module, the complex background removal and mask extraction of the multimodal response data include:

[0042] Based on the text prompts, the Grounding DINO model is used to process the multimodal response data to generate labeled target detection boxes, and negative masks are generated to remove interference at the pixel level. Then, the target detection boxes with interference removed are input into the SAM model to generate high-precision vulvar and ear mask regions.

[0043] In the multimodal feature quantization module, constructing visual features includes:

[0044] A fully associative tracking algorithm was used to calculate the standing time and pen-arching frequency of sows after being stimulated by estrus, and to calculate the RGB color histogram by combining the mask region to obtain visual features;

[0045] Constructing thermal infrared features includes:

[0046] The thermal infrared image of the body surface infrared temperature data is registered with the visible light image, and the highest temperature, average temperature and the difference between the temperature of the vulva mask area and the ear root temperature are extracted to obtain the thermal infrared features.

[0047] Constructing acoustic features includes:

[0048] The audio response data is denoised, the Mel frequency cepstral coefficients are extracted, and a convolutional network is used to identify high-frequency short humming patterns to obtain acoustic features.

[0049] Furthermore, the attention-based large model diagnostic subsystem includes a multimodal cross-attention fusion module, a LoRA large model, and a decision diagnosis and interactive closed-loop execution module;

[0050] The multimodal cross-attention fusion module is used to fuse multidimensional response features to form fused features;

[0051] The LoRA large model is fine-tuned based on the pig estrus induction multimodal instruction fine-tuning dataset and deployed on the edge computing unit carried by the estrus induction and estrus detection robot. It is used to process the fused features to obtain a diagnostic report that includes confidence analysis and mating suggestions.

[0052] The decision diagnosis and interactive closed-loop execution module executes the decision feedback loop of the ejaculation process based on the diagnosis report;

[0053] When estrus is diagnosed, the estrus-inducing robot is activated to physically mark the sow and upload the diagnostic log.

[0054] When a sow is diagnosed as not in estrus and a single interaction does not reach the preset time threshold, the parameters of the bionic estrus induction mechanism are fine-tuned. When the preset time threshold has been reached and the sow is still not in estrus, the current interaction ends and the mobile adaptation chassis is driven to move to the next sow pen.

[0055] Among them, the fine-tuning parameters of the biomimetic arousal mechanism include the control parameters for generating the modulation control sequence, the audio modulation parameters, and the release frequency of the odor pulse.

[0056] Furthermore, in the multimodal cross-attention fusion module, multidimensional response features are fused to form fused features, including:

[0057] Flatten the visual features into a query tensor The thermal infrared features and audio features are spliced ​​together to form a key. Sum Tensors, through calculation The feature weights of thermal infrared features and audio features are determined, and then the visual features, thermal infrared features and audio features are spliced ​​together to obtain the fused features.

[0058] A working method for an intelligent biomimetic arousal and relationship-detecting robot system includes:

[0059] S100, Global Initialization: Wake up the system and perform sensor calibration, and perform a panoramic scan of the pigsty during the movement of the mobile adapter chassis;

[0060] S200, Target Locking: Determine the face and ear mask of the target sow, continuously calculate the real-time distance from the scent nozzle of the estrus detection robot to the sow's snout, and activate follow-up tracking;

[0061] S300, Bionic stimulus: Responds to the target mask exceeding a set consecutive frames when the target sow detection box confidence is greater than a set threshold.

[0062] A four-dimensional hyperchaotic system is activated to generate a modulation control sequence and dynamically modulates pre-recorded pig audio samples, thereby transmitting anti-habitation directional estrus-inducing audio.

[0063] The real-time distance is input into the inverse model trained based on the distance compensation relationship, and the corresponding duty cycle PWM control signal is output to control the solenoid valve of the odor nozzle to release odor irritants and perform micro-distance approach-backward probing action.

[0064] S400, Feature Extraction: During the observation window of continuous excitation, visual features, thermal infrared features, and acoustic features are extracted and fused and compensated to generate fused features;

[0065] S500, Estrus Diagnosis: After the observation window period ends, the fused features and time indicators are input into the LoRA large model. Estrus diagnosis is performed based on the duration of static standing with both ears, the duration of active contact with the estrus-inducing robot, and the consistency of multimodal features, and a diagnostic report is generated.

[0066] In response to the diagnosis that the sow is not in estrus, the control parameters of the modulation control sequence for generating the variable frequency sound wave, the audio modulation parameters, and the release frequency of the odor pulse are finely adjusted, and the process returns to S300.

[0067] This invention, through the organic combination of multimodal fusion technology and a habit-resistant biomimetic stimulus mechanism, has produced the following significant technological advancements and beneficial effects:

[0068] (1) Significantly improves the estrus detection rate and completely solves the problem of missed detection of "silent estrus":

[0069] Unlike traditional passive monitoring (relying solely on camera observation of natural behavior), this invention utilizes multi-dimensional active stimulation composed of variable-frequency sound waves and precise PWM odor to simulate the biofield effect of a boar in the presence of the sow. This "active interaction" mechanism effectively activates the hypothalamus-pituitary-ovarian axis in sows, inducing standing reflexes in sows in early estrus or with mild symptoms. Simulation results and comparative tests show that this invention increases the estrus detection rate from 80%-90% with traditional visual methods to over 98%, particularly significantly improving the detection rate for silent estrus sows from less than 50% to over 90%, greatly reducing economic losses caused by missed matings.

[0070] (2) Fundamentally address the diminishing effect caused by animalistic habits:

[0071] Existing bionic robots, relying on fixed recordings and continuous spraying, often fail after 1-2 weeks of use due to sows developing adaptation. This invention introduces a chaotic frequency-conversion acoustic algorithm that generates sound waves with mathematically infinite non-repetitiveness in both the time and frequency domains. Combined with PWM-controlled pulsed odor gradients, this breaks down the sows' sensory adaptation at a neurophysiological level. This design ensures the device maintains high arousal levels for the sows throughout its entire lifespan, solving the long-term effectiveness problem and transforming the robot from a short-term effective "toy" into a long-term reliable production tool.

[0072] (3) To achieve precise quantification and conservation of sex pheromones, and to prevent olfactory desensitization:

[0073] Traditional spray systems cannot control the dosage, wasting expensive hormone reagents and easily causing odor pollution in pigsties. The PWM duty cycle control technology of this invention dynamically adjusts the spray volume according to distance, achieving "on-demand supply." Comparative tests show that, while achieving the same stimulation effect, this system can save approximately 60% of the amount of pheromone reagent used. Simultaneously, the pulsed release method avoids olfactory receptor saturation in sows due to excessively high ambient odor concentrations, ensuring the effectiveness of each stimulation.

[0074] (4) Enhance robustness and anti-interference capability in complex environments:

[0075] Traditional vision systems are susceptible to interference from factors such as lighting conditions in pigsties, fence occlusion, and background pollution from feces. This invention employs Grounded-SAM for semantic-level instance segmentation. By inputting negative textual cues (such as "iron fence" and "background pollution"), the model is guided to generate a negative mask, thereby achieving accurate removal of physical occlusion and background noise. More importantly, the multimodal cross-attention mechanism eliminates reliance on a single metric. Even under conditions of extremely poor lighting leading to visual failure, thermal infrared (external temperature) and acoustic (feeble) features can still achieve accurate judgment through the dominant role of attention weights. This complementary redundancy design significantly reduces the false alarm rate to below 2% under various harsh conditions.

[0076] (5) Real-time reasoning and interpretability diagnosis at the edge:

[0077] Utilizing LoRA fine-tuning technology, this invention successfully deployed a massive multimodal model on the edge computing unit of a mobile robot, achieving millisecond-level offline real-time response and solving the data transmission latency problem caused by poor network environments in pig farms. The detection time for a single sow was reduced to 10-20 seconds. Furthermore, the model outputs more than just cold "yes / no" results; it includes natural language reports containing symptom descriptions, confidence levels, and mating time recommendations (e.g., "Strong standing reflex detected, vulvar redness and swelling subsided, indicating peak estrus; immediate mating recommended"). This significantly improves the system's usability and its guidance value for frontline personnel. Attached Figure Description

[0078] Figure 1 The diagram shows the structure of the intelligent bionic arousal and estrus detection robot system provided by this invention.

[0079] Figure 2 The working logic diagram of the intelligent bionic arousal and estrus detection robot system provided by the present invention. Detailed Implementation

[0080] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0081] Example 1:

[0082] like Figure 1 As shown, an intelligent bionic arousal and flirtation detection robot system includes:

[0083] The autonomous mobile inspection platform serves as the physical carrier of the seduction and seduction robot;

[0084] The anti-habit biomimetic excitation subsystem is used to generate modulation control sequences and dynamically modulate pre-recorded pig audio samples to output anti-habit biomimetic estrus-inducing audio, and combine it with odor pulses to trigger a biomimetic estrus-inducing mechanism.

[0085] The multimodal perception and feature extraction subsystem is used to perceive the physiological and behavioral response data of sows after being stimulated by the biomimetic estrus mechanism, and convert them into multidimensional response features.

[0086] The large model diagnostic subsystem based on cross-attention is used to perform cross-attention fusion on multi-dimensional response features, obtain estrus diagnosis results based on the fused features, and drive the system to perform closed-loop execution.

[0087] The intelligent biomimetic estrus-inducing robot system provided in this invention adopts a highly integrated and modular design concept. Physically, it simulates the form of a boar to generate visual stimulation. Functionally, it integrates four core capabilities: autonomous navigation, active stimulation, multi-dimensional perception, and edge computing. Based on this, the system is functionally divided into four core subsystems: an autonomous mobile inspection platform subsystem, a behavior-resistant biomimetic stimulation subsystem, a multi-modal perception and feature extraction subsystem, and a large-model diagnostic subsystem based on cross-attention.

[0088] In this embodiment of the invention, the autonomous mobile inspection platform is the physical carrier of the robot, providing complete support from physical spatial anchoring to triggering erogenous zones. Specifically, it includes:

[0089] Functional exterior structure, mobile-adaptive chassis, navigation module, and identity binding module;

[0090] The exterior functional structure adopts a static biomimetic boar shell and is equipped with an acoustic waveguide, a hidden odor nozzle and a ranging sensor;

[0091] The mobile adaptable chassis, equipped with a functional structure, allows for movement within the pigsty.

[0092] The navigation module uses laser SLAM to build a map and combines it with UWB tags to guide the movement and stopping of the estrus detection robot in the pigsty;

[0093] After the estrus detection robot stops, the identity binding module collects multimodal response data by reading the sow's ear tag with RFID and binds it to the individual's identity.

[0094] In this embodiment, the functionalized external structure adopts a static biomimetic boar shell covered with pink silicone material. Acoustic waveguides are integrated on both sides of the head to project arousal audio in a directional manner, and a hidden odor nozzle and ranging sensor are built into the nose; this design structure provides a multi-dimensional signal field generation interface for subsequent active arousal.

[0095] In this embodiment, for the narrow passages of pigsties, which are typically 0.8-1.2 meters wide, and the slippery ground with residual manure, the mobile adapter chassis adopts a Mecanum wheel omnidirectional movement structure and an independent suspension shock absorption system to ensure that the platform operates smoothly and can turn on the spot.

[0096] In this embodiment, the navigation module uses laser SLAM to build a map and combines it with UWB tags to assist in positioning to eliminate the cumulative error caused by the pigpen corridor effect, ensuring that the robot can navigate accurately and stop directly in front of the designated sow pen.

[0097] In this embodiment, for the identity binding module, after the estrus detection machine stops, the side RFID module reads the sow's ear tag and accurately binds the collected multimodal response data with the individual's identity, and then triggers the anti-habit bionic stimulus subsystem.

[0098] In this embodiment of the invention, the anti-habit biomimetic excitation subsystem solves the problem of sow's perception and adaptation by simulating the biological characteristics of boars and introducing a chaotic random mutation mechanism; it includes an anti-habit hyperchaotic sound field generation module deployed in the edge computing unit of the estrus-inducing and estrus-detecting robot and a closed-loop control loop based on physical law constraints.

[0099] In this embodiment, the anti-habit hyperchaotic sound field generation module is equipped with a hyperchaotic dynamics model. It generates a modulation control sequence through a four-dimensional hyperchaotic system and applies the modulation control sequence to pre-recorded pig audio samples to output anti-habit biomimetic estrus-inducing audio with high biosimilarity and unpredictability in the time and frequency domains.

[0100] In this embodiment, a closed-loop control loop based on physical constraints uses the real-time distance between the odor nozzle of the estrus detection robot and the sow's snout as the input of the inverse model. The required odor stimulation intensity to reach the effective stimulation threshold of the sow's snout is calculated, and a PWM control signal with a corresponding duty cycle is dynamically output based on the change in real-time distance to drive the solenoid valve of the odor nozzle that releases hormones, thereby maintaining the effective odor stimulation intensity of the target area in the pig house environment.

[0101] In this embodiment, the underlying hardware of the anti-habitual hyperchaotic sound field generation module consists of a digital signal processor (DSP), a digital-to-analog converter (DAC), and a voltage-controlled oscillator (VCO) hardware circuit built into the edge computing unit.

[0102] To overcome the shortcomings of traditional low-dimensional chaos, this invention employs an improved four-dimensional hyperchaotic system to generate the control signal. This system has two positive Lyapunov exponents, ensuring that the separation of the output signal's trajectory in phase space increases exponentially. Based on this, the dynamic equations for generating the control signal in the habit-resistant hyperchaotic sound field generation module are as follows:

[0103]

[0104] In the formula, For the state variables of a four-dimensional hyperchaotic system; Representing state variables respectively Rate of change over time; , , , These are system control parameters or coupling parameters used to adjust the coupling strength and nonlinear evolution characteristics between various state variables of the system. The bias term is used to drive the continuous evolution of the fourth-dimensional state variable; wherein, the state variable is mapped to form an audio modulation control quantity, which is used to dynamically modulate one or more of the fundamental frequency parameters, amplitude envelope parameters, formant parameters and rhythm parameters of the pre-recorded pig audio samples.

[0105] In this embodiment, pre-recorded audio samples of pigs are used as basic audio inputs to the audio modulation module; the DSP runs the four-dimensional hyperchaotic dynamic equations to generate a random control sequence, which is converted by the DAC to form a modulation control sequence or analog control voltage, and input to the audio modulation module to dynamically modulate the basic audio samples, thereby outputting anti-habit biomimetic estrus-inducing audio.

[0106] In this invention, the aforementioned anti-habit hyperchaotic sound field generation module utilizes chaotic mapping and stochastic process algorithms to dynamically modulate one or more of the fundamental frequency parameters, formant parameters, amplitude envelope parameters, and rhythm parameters of pre-recorded pig audio samples, outputting anti-habit biomimetic estrus-inducing audio with high biosimilarity and unpredictability in both the time and frequency domains. This breaks the auditory adaptation of sows at the neural mechanism level, ensuring continuous and efficient auditory arousal. The method uses mathematical means to ensure that the output audio has long-term non-repetitive characteristics.

[0107] In this embodiment, for the aforementioned set of dynamic equations that generate control signals, the complete and closed-loop flow path of the relevant data is as follows:

[0108] Input settings: "Pre-recorded pig audio samples are used as the basic audio input"; in the initial stage of the dynamic equation iteration, "if it is the first excitation, the initial chaotic parameters are randomly generated" as the operation trigger input.

[0109] Output side settings and physical mapping: The hardware destination of the solution results is that the DSP runs the four-dimensional hyperchaotic dynamic equations to generate a random control sequence, and the random control sequence is converted by the DAC to form a modulation control signal or analog control voltage.

[0110] Physical target control: The target of the control voltage is not "used to dynamically modulate one or more of the fundamental frequency parameters, amplitude envelope parameters, formant parameters, and rhythm parameters of pre-recorded pig audio samples".

[0111] Furthermore, during the operation and fine-tuning phases, a feedback-based closed-loop adjustment mechanism, namely a "system-triggered parameter fine-tuning mechanism," automatically adjusts the control parameters of the four-dimensional hyperchaotic system to change the audio modulation characteristics. Thus, the system control parameters in the equation set are variables that are automatically adjusted based on multimodal feedback results during system operation. This parameter setting method embodies the dynamic adjustment characteristics that achieve "anti-habitualization."

[0112] Therefore, the above-mentioned set of dynamic equations provided by the present invention constitutes a complete implementation link from the digital domain to the analog domain and finally to the acoustic physics domain.

[0113] In this embodiment, during the functional implementation of the closed-loop control loop based on physical constraints, to reliably acquire spatial distance in complex pigsties, an RGB-D depth camera is used to identify the bounding box of the sow's head, and the average Z-axis depth value is extracted from the depth map as the real-time distance. To adapt to the characteristics of odor propagation varying with distance in complex breeding environments, this invention constructs an inverse model trained based on a distance compensation relationship, and uses the compensation relationship between odor stimulation intensity and distance as a training constraint to improve the stability and rationality of the model output. The edge computing unit deploys the inverse model, uses the real-time distance as input, calculates the odor stimulation intensity required to reach the effective stimulation threshold of the sow's snout, and dynamically outputs a PWM control signal with a corresponding duty cycle to drive the solenoid valve, thereby maintaining the effective odor stimulation intensity of the target area in dynamic space.

[0114] In this invention, the closed-loop control circuit based on physical constraints incorporates a pulse width modulation (PWM) algorithm to control the solenoid valve, achieving millisecond-level precise control of the pheromone release concentration, frequency, and duration by adjusting the duty cycle. A pulsed odor gradient is constructed, utilizing the sensitivity of olfactory neurons to the rate of concentration change (dC / dt) to ensure both instantaneous stimulation intensity and prevent receptor saturation. Simultaneously, a distance feedback algorithm significantly reduces the cost of expensive hormone reagents.

[0115] In this embodiment of the invention, the biomimetic erotic induction mechanism triggered by the aforementioned anti-habit biomimetic stimulation subsystem solves the problem that existing biomimetic devices, due to their monotonous and repetitive sound and behavioral patterns, cause the sow's brainstem reticular activating system (RAS) to recognize the periodicity of signals, thus classifying them as invalid background noise and performing neural adaptive filtering. This biological mechanism leads to sows becoming sluggish or even ignoring stimuli. To overcome this biological barrier of the RAS, this invention introduces deterministic chaos at the signal source, utilizing its long-term unpredictability to generate a non-periodic modulation control sequence adapted to the biological brain, and outputs anti-habit biomimetic erotic induction audio, thereby solving the problem of significantly reduced standing reflex induction rate, which leads to the device becoming an ineffective decoration after long-term deployment.

[0116] Furthermore, to address the issue of odor field construction failure caused by neglecting Fick's diffusion law in traditional spray systems, and the problem that traditional constant spraying leads to excessively high concentrations (desensitization) at close range or excessively low concentrations (ineffectiveness) at distant ranges, this invention establishes a real-time "distance-concentration" compensation model based on physical principles (odor molecule concentration decreases non-linearly with increasing propagation distance). This model ensures instantaneous stimulation intensity while preventing receptors from losing sensitivity due to continuous exposure to high concentrations.

[0117] In this embodiment, when the biomimetic estrus induction mechanism (biomimetic estrus induction audio and odor pulse) of the anti-habit biomimetic stimulation subsystem is triggered, the multimodal perception and feature extraction subsystem can be started simultaneously, forming a closed loop of "active induction-synchronous perception". This subsystem is responsible for comprehensively perceiving the physiological data and behavioral responses of the sow after being stimulated by estrus, and transforming the raw data into high-level semantic features.

[0118] The multimodal perception and feature extraction subsystem in this embodiment of the invention includes a multimodal response data synchronous acquisition module and a multimodal feature quantization module;

[0119] The multimodal response data collected by the multimodal response data synchronous acquisition module includes visible light image sequences, body surface infrared temperature data, audio response data, and three-dimensional depth data for sows.

[0120] The multimodal feature quantization module constructs multidimensional response features, including visual features, thermal infrared features, and acoustic features, after complex background removal and mask extraction of the collected multimodal response data.

[0121] In this embodiment, upon receiving the erogenous zone trigger command, the system uses a hardware clock synchronization trigger mechanism to collect the following multimodal response data via the multimodal response data synchronization acquisition module:

[0122] Visible light image sequence: A high-definition visible light camera (4K resolution, 60fps, equipped with a servo gimbal) was used to synchronously acquire a sequence of images of the sow’s subtle external movements (such as ear tremors, eye contact, and external features such as vulvar redness and swelling, and mucus).

[0123] Infrared body surface temperature data: Non-contact measurement was performed using an infrared thermal imager (resolution 640x512, thermal sensitivity NETD<40mK) to collect absolute values ​​of body surface temperature and thermal distribution arrays of key areas such as the vulva, ear base, and eyes of sows.

[0124] Audio response data: Using a microphone array (supporting sound source localization and echo cancellation), audio feedback data of target sows after being induced is picked up and recorded in noisy pigsty background noise.

[0125] 3D depth data: RGB-D depth camera is used to acquire 3D point cloud data of sows to help determine accurate standing posture and volume changes of vulvar swelling.

[0126] In this embodiment, after acquiring the original multimodal data, the multimodal feature quantization module performs complex background removal and mask extraction on the multimodal response data, including:

[0127] Based on the text prompts, the Grounding DINO model is used to process the multimodal response data to generate labeled target detection boxes. Negative masks are generated to remove interfering objects at the pixel level. Then, the target detection boxes with removed interfering objects are input into the SAM model to generate high-precision vulvar and ear mask regions.

[0128] Specifically, to address the issue of complex background interference in pigsties, such as fence obstruction and manure, the Grounding DINO model is used. Input text prompts (e.g., "sow vulva," "pig ear," "iron fence," "background manure") generate labeled detection boxes. For interfering objects like fences, a negative mask is generated for pixel-level removal. The target detection boxes are then input into the SAM (Segment Anything Model) to generate a denoised, high-precision mask region for the vulva and ears.

[0129] In this embodiment, in the multimodal feature quantization module:

[0130] Constructing visual features include:

[0131] A fully associative tracking algorithm is employed to calculate the standing time and pen-arching frequency of sows after being stimulated by estrus, and to obtain visual features by calculating an RGB color histogram in conjunction with the masked region. Preferably, ByteTrack's fully associative tracking algorithm is used to maintain the continuity of the target pig ID in the presence of physical occlusion, and to accurately calculate the "standing time" and "pen-arching frequency" of sows after being stimulated by estrus.

[0132] Constructing thermal infrared features include:

[0133] Register the thermal infrared image of the body surface infrared temperature data with the visible light image to extract the highest temperature in the vulvar mask area. Average temperature And the temperature difference with the base of the ear. Thermal infrared characteristics were obtained.

[0134] Constructing acoustic features include:

[0135] The audio response data is denoised, the Mel frequency cepstral coefficients are extracted, and a convolutional network is used to identify high-frequency short humming patterns to obtain acoustic features.

[0136] In this embodiment of the invention, the above-mentioned multimodal perception and feature extraction subsystem solves the problem that existing systems rely only on single visual or body temperature data, or only perform simple voting fusion, lacking effective alignment and complementary fusion of heterogeneous data such as video, infrared, and audio. Under complex working conditions such as changes in lighting, object occlusion, and background interference, single modality is prone to failure, resulting in low detection accuracy and high false alarm rate.

[0137] In this embodiment of the invention, the large model diagnostic subsystem based on cross-attention serves as the "brain" of the robot, responsible for receiving the previously extracted multimodal features, performing intelligent reasoning and diagnosis, and outputting decision instructions.

[0138] The large model diagnostic subsystem based on cross-attention in this embodiment of the invention includes a multimodal cross-attention fusion module, a LoRA large model, and a decision diagnosis and interactive closed-loop execution module;

[0139] The multimodal cross-attention fusion module is used to fuse multidimensional response features to form fused features;

[0140] The LoRA large model is fine-tuned based on a dataset of pig estrus induction multimodal commands and deployed on the edge computing unit carried by the estrus induction and estrus detection robot. It is used to process fused features to obtain a diagnostic report that includes confidence analysis and mating suggestions.

[0141] The decision diagnosis and interactive closed-loop execution module executes a decision feedback loop based on the diagnostic report to induce the ejaculation process.

[0142] When estrus is diagnosed, the estrus-inducing robot is activated to physically mark the sow and upload the diagnostic log.

[0143] When a sow is diagnosed as not in estrus and a single interaction does not reach the preset time threshold, the parameters of the bionic estrus induction mechanism are fine-tuned. When the preset time threshold has been reached and the sow is still not in estrus, the current interaction ends and the mobile adaptation chassis is driven to move to the next sow pen.

[0144] Among them, the fine-tuning parameters of the biomimetic arousal mechanism include the control parameters for generating the modulation control sequence, the audio modulation parameters, and the release frequency of the odor pulse.

[0145] In this embodiment, to address the issue of decreased confidence in feature acquisition due to physical occlusion (such as large iron bars obstructing the standing movements of pigs) or sudden changes in lighting in actual pigsties, this invention designs a proprietary multimodal cross-attention fusion mechanism. In the multimodal cross-attention fusion module, multidimensional response features are fused to form fused features, including:

[0146] Flatten the visual features into a query tensor The thermal infrared features and audio features are spliced ​​together to form a key. Sum Tensors, through calculation The feature weights of thermal infrared features and audio features are determined, and then the visual features, thermal infrared features and audio features are spliced ​​together to obtain the fused features.

[0147] Specifically, the visual features extracted from the preceding sequence that contain temporal action information are... (e.g., standing time) Flattening is defined as query tensor (Q), which is minimally affected by environmental physical obstruction of thermal infrared features. (Vulvar temperature difference) and audio characteristics (The spectrum of estrus moans) is spliced ​​as a key (K) and value (V) tensor. This is achieved through computation... When visual signals are severely obstructed, When the modulus decreases, the attention mechanism automatically searches for and amplifies the feature weights of thermal infrared (sudden temperature rise) or audio (specific calls) that are highly synchronized with the current action on the time axis. This mechanism uses acoustic and thermal features to semantically compensate for incomplete visual features, greatly improving the robustness of discrimination under complex arousal conditions.

[0148] In this invention, the aforementioned multimodal cross-attention fusion module utilizes the cross-attention mechanism in the Transformer architecture to deeply fuse visible light images (behavior / appearance), infrared thermal images (body temperature distribution), and audio signals (vocal features). This mechanism allows the model to dynamically capture strongly correlated features between modalities (such as the temporal synchronization between "ear-erecting action" and "high-frequency humming"), and even when a certain modality is missing (e.g., visual occlusion) or interfered with, accurate determination can be achieved through weight compensation from other modalities.

[0149] In this embodiment, addressing the pain points of traditional pig farms with no or weak network environments and limited computing power at the edge of mobile robots (such as limited video memory on the embedded computing board), this invention does not rely on cloud inference but instead adopts a lightweight fine-tuning strategy based on LoRA (Low-Rank Adaptation).

[0150] Specifically, a dedicated "pig estrus induction multimodal command fine-tuning dataset" was constructed, containing real sensor data and natural language diagnostic labels from breeding experts for fine-tuning the LoRA large-scale model. During the fine-tuning phase, the pre-trained weight matrix of the vision-language large-scale model (such as Qwen2-VL) was frozen. Low-rank factorization matrices are injected and trained only at the Transformer attention layer. and Because only a very small number of low-rank parameters are optimized, memory usage drops dramatically. The trained weights are quantized with Int8 and seamlessly deployed on the edge computing unit on the robot, enabling millisecond-level real-time inference in offline mode, and finally outputting a natural language diagnostic report containing confidence analysis and breeding suggestions.

[0151] This invention utilizes Low-Rank Adaptation (LoRA) technology for efficient fine-tuning of a large vision-language model. By freezing pre-trained weights and training only the low-rank matrix, it significantly reduces memory usage and computational requirements while maintaining diagnostic performance. This allows for smooth deployment on the robot's onboard computing platform, achieving lightweight deployment and real-time inference without sacrificing accuracy, and generating interpretable diagnostic results. This addresses the problems of traditional high-precision multimodal deep learning models, which, due to their massive computational demands, are difficult to run in real-time on the limited embedded computing units of mobile robots, and are also ill-suited to adapting to different pigsty environments.

[0152] In this embodiment, in the decision diagnosis and interactive closed-loop execution module, after the large model outputs a diagnostic report, this module will drive the system into the closed-loop execution stage to fill the decision feedback loop in the active inducement process:

[0153] Positive response execution: If estrus is diagnosed, this subsystem sends a command via the bus to drive the inkjet device at the tail of the robot to physically mark the back of the sow and simultaneously upload the diagnostic log.

[0154] Negative Response Closed Loop (Anti-Habitual Retry): If the current diagnosis is negative and the single interaction does not reach the preset time threshold (e.g., 60 seconds), this subsystem will trigger the parameter fine-tuning mechanism, send parameter tuning instructions to the bionic excitation subsystem, forcibly adjust the control parameters of the four-dimensional hyperchaotic system to generate a modulation control sequence with higher novelty, output a bionic induced arousal audio with stronger anti-adaptive properties, increase the PWM odor pulse frequency, and perform enhanced secondary compound induced arousal; if there is still no positive response after reaching the threshold, the current interaction will end, the baseline library will be updated, and the chassis will be driven to the next column.

[0155] This invention realizes proactive interactive inspection and semantic-level diagnostic reports. In the process, a mobile robot with autonomous navigation function is used to simulate the boar's pen patrol behavior and actively induce the standing reflex through multi-dimensional sensory stimulation. At the same time, combined with Large Language Model (LLM) technology, an interpretable natural language diagnostic report is generated, which includes symptom description, confidence analysis and mating suggestions, to assist human decision-making and realize a closed loop from "data collection" to "decision support".

[0156] Example 2:

[0157] This embodiment is a further limitation based on Embodiment 1. Its purpose is to describe the working method of the intelligent bionic arousal and estrus detection robot system. Other parts not mentioned refer to the description in Embodiment 1 or the prior art.

[0158] refer to Figure 2 A working method for an intelligent biomimetic arousal and relationship-detecting robot system includes:

[0159] S100, Global Initialization: Wake up the system and perform sensor calibration, and perform a panoramic scan of the pigsty during the movement of the mobile adapter chassis;

[0160] Specifically, when the set time frame is reached or a remote inspection command is received, the omnidirectional chassis SLAM navigation algorithm is activated, which integrates point cloud and RGB vision to avoid obstacles, and travels to the target large group inspection area according to the preset map, and the sensor gimbal performs a panoramic scan.

[0161] S200, Target Locking: Determine the face and ear mask of the target sow, continuously calculate the real-time distance from the scent nozzle of the estrus detection robot to the sow's snout, and activate follow-up tracking;

[0162] Specifically, when the mobile adapter chassis is stationary and the gimbal enters the target search field of view, the edge-end Grounded-SAM vision base is activated to lock onto the independent sow target and output a high-precision pixel mask of the pig's face and ears; furthermore, the binocular depth camera continuously calculates the absolute spatial straight-line distance from the robot's nozzle to the sow's snout, and the gimbal starts follow-up tracking.

[0163] S300, Bionic stimulus: Responds to the target mask exceeding a set consecutive frames when the target sow detection box confidence is greater than a set threshold.

[0164] The auditory channel activates a four-dimensional hyperchaotic system to generate a modulation control sequence and dynamically modulates pre-recorded audio samples of pigs, thereby emitting anti-habitation directional erotic induction audio.

[0165] The visual channel inputs real-time distance into the inverse model trained based on distance compensation relationship, and outputs a PWM control signal with corresponding duty cycle to control the solenoid valve of the odor nozzle to release odor stimulants and perform micro-distance approach-backward probing action.

[0166] S400, Feature Extraction: During the observation window of continuous excitation, visual features, thermal infrared features, and acoustic features are extracted and fused and compensated to generate fused features;

[0167] Specifically, when non-contact active stimulation is performed continuously for more than 10 seconds, within the observation window (3 minutes) of continuous stimulation, the visual channel extracts the rate of change of the binaural posture angle and records the "total duration of binaural static standing" and the "total duration of active close contact"; the infrared array extracts the thermal anomaly of the vulva; the sound pickup array extracts the response audio; further, the cross-attention network is triggered to fuse and compensate the three features to generate a high-dimensional feature matrix.

[0168] S500, Estrus Diagnosis: After the observation window period ends, the fused features and time indicators are input into the LoRA large model. Estrus diagnosis is performed based on the duration of static standing with both ears, the duration of active contact with the estrus-inducing robot, and the consistency of multimodal features, and a diagnostic report is generated.

[0169] Specifically, at the end of the observation window (default 3 minutes), the high-dimensional feature matrix and time indicators are input into the LoRA large model. If the cumulative duration of the sow remaining in a static state reaches 41.3 seconds, and the duration of active contact with the device reaches 29.7 seconds, and the multimodal feature consistency is high, then a "back-pressure static reflex" is equivalently determined. A natural language diagnostic report is then generated locally and asynchronously transmitted to the control panel, from where it is transferred to the next sow.

[0170] In this embodiment, during estrus diagnosis, if the result is positive (estrus), the sow in estrus is physically marked (inkjet printing).

[0171] If the diagnosis is negative (no standing reaction): First determine whether the current single-cell interaction duration has reached the preset threshold (e.g., 60 seconds).

[0172] If the threshold is not reached: the system triggers a parameter fine-tuning mechanism to automatically adjust the control parameters of the four-dimensional hyperchaotic system. To change the frequency characteristics of the sound wave and simultaneously increase the frequency of the PWM odor pulse, the signal is returned to S300 for enhanced secondary excitation.

[0173] If the threshold is reached but the result is still negative: it is determined that the sow is not currently in estrus. After recording the data and updating the historical baseline database, the robot moves to the next pen.

[0174] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0175] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. An intelligent biomimetic arousal and relationship-detecting robot system, characterized in that, include: The autonomous mobile inspection platform serves as the physical carrier of the seduction and seduction robot; The anti-habit biomimetic excitation subsystem is used to generate modulation control sequences and dynamically modulate pre-recorded pig audio samples to output anti-habit biomimetic estrus-inducing audio, and combine it with odor pulses to trigger a biomimetic estrus-inducing mechanism. The multimodal perception and feature extraction subsystem is used to perceive the physiological and behavioral response data of sows after being stimulated by the biomimetic estrus mechanism, and convert them into multidimensional response features. The large model diagnostic subsystem based on cross-attention is used to perform cross-attention fusion on multi-dimensional response features, obtain estrus diagnosis results based on the fused features, and drive the system to perform closed-loop execution.

2. The intelligent bionic arousal-inducing and arousal-detecting robot system according to claim 1, characterized in that, The autonomous mobile inspection platform includes a functional structure, a mobile-adaptive chassis, a navigation module, and an identity binding module. The functionalized external structure adopts a static biomimetic boar shell and is equipped with an acoustic waveguide, a hidden odor nozzle, and a ranging sensor. The mobile adapter chassis, equipped with a functional structure, moves within the pigsty. The navigation module uses laser SLAM to build a map and combines it with UWB tags to navigate the movement and stopping of the estrus detection robot in the pigsty; After the estrus detection robot stops, the identity binding module collects multimodal response data by reading the sow's ear tag with RFID and binds it to the individual's identity.

3. The intelligent bionic arousal and estrus detection robot system according to claim 1, characterized in that, The anti-habit bionic excitation subsystem includes an anti-habit hyperchaotic sound field generation module and a closed-loop control loop based on physical law constraints, deployed in the edge computing unit of the erotic detection robot. The anti-habit hyperchaotic sound field generation module is equipped with a hyperchaotic dynamics model. It generates a modulation control sequence through a four-dimensional hyperchaotic system and applies the modulation control sequence to pre-recorded pig audio samples to output anti-habit biomimetic estrus-inducing audio with high biosimilarity and unpredictability in the time and frequency domains. The closed-loop control circuit based on physical constraints uses the real-time distance between the odor nozzle of the estrus detection robot and the sow's snout as the input of the inverse model. It calculates the odor stimulation intensity required to reach the effective stimulation threshold of the sow's snout and dynamically outputs a PWM control signal with a corresponding duty cycle based on the change in real-time distance to drive the solenoid valve of the odor nozzle that releases hormones, thereby maintaining the effective odor stimulation intensity of the target area in the pig house environment.

4. The intelligent bionic arousal and estrus detection robot system according to claim 3, characterized in that, In the anti-habitual hyperchaotic sound field generation module, the dynamic equations for generating the control signal are as follows: In the formula, For the state variables of a four-dimensional hyperchaotic system, Representing state variables respectively Rate of change over time; , , , These are system control parameters or coupling parameters used to adjust the coupling strength and nonlinear evolution characteristics between various state variables of the system. The bias term is used to drive the continuous evolution of the fourth-dimensional state variable; wherein, the state variable is mapped to form an audio modulation control quantity, which is used to dynamically modulate one or more of the fundamental frequency parameters, amplitude envelope parameters, formant parameters and rhythm parameters of the pre-recorded pig audio samples.

5. The intelligent bionic arousal-inducing and arousal-detecting robot system according to claim 3, characterized in that, During pre-training, the inverse model uses the compensation relationship between odor stimulus intensity and distance as a training constraint to improve the stability and rationality of the inverse model's output.

6. The intelligent bionic arousal and estrus detection robot system according to claim 1, characterized in that, The multimodal perception and feature extraction subsystem includes a multimodal response data synchronous acquisition module and a multimodal feature quantization module; The multimodal response data collected by the multimodal response data synchronous acquisition module includes visible light image sequences, surface infrared temperature data, audio response data, and three-dimensional depth data for sows. The multimodal feature quantization module constructs multidimensional response features, including visual features, thermal infrared features, and acoustic features, after complex background removal and mask extraction of the collected multimodal response data.

7. The intelligent bionic arousal and estrus detection robot system according to claim 6, characterized in that, In the multimodal feature quantization module, complex background removal and mask extraction of multimodal response data include: Based on the text prompts, the Grounding DINO model is used to process the multimodal response data to generate labeled target detection boxes, and negative masks are generated to remove interference at the pixel level. Then, the target detection boxes with interference removed are input into the SAM model to generate high-precision vulvar and ear mask regions. In the multimodal feature quantization module, constructing visual features includes: A fully associative tracking algorithm was used to calculate the standing time and pen-arching frequency of sows after being stimulated by estrus, and to calculate the RGB color histogram by combining the mask region to obtain visual features; Constructing thermal infrared features includes: The thermal infrared image of the body surface infrared temperature data is registered with the visible light image, and the highest temperature, average temperature and the difference between the temperature of the vulva mask area and the ear root temperature are extracted to obtain the thermal infrared features. Constructing acoustic features includes: The audio response data is denoised, the Mel frequency cepstral coefficients are extracted, and a convolutional network is used to identify high-frequency short humming patterns to obtain acoustic features.

8. The intelligent bionic arousal and arousal detection robot system according to claim 1, characterized in that, The attention-based large model diagnostic subsystem includes a multimodal cross-attention fusion module, a LoRA large model, and a decision diagnosis and interactive closed-loop execution module. The multimodal cross-attention fusion module is used to fuse multidimensional response features to form fused features; The LoRA large model is fine-tuned based on the pig estrus induction multimodal instruction fine-tuning dataset and deployed on the edge computing unit carried by the estrus induction and estrus detection robot. It is used to process the fused features to obtain a diagnostic report that includes confidence analysis and mating suggestions. The decision diagnosis and interactive closed-loop execution module executes the decision feedback loop of the ejaculation process based on the diagnosis report; When estrus is diagnosed, the estrus-inducing robot is activated to physically mark the sow and upload the diagnostic log. When a sow is diagnosed as not in estrus and a single interaction does not reach the preset time threshold, the parameters of the bionic estrus induction mechanism are fine-tuned. When the preset time threshold has been reached and the sow is still not in estrus, the current interaction ends and the mobile adaptation chassis is driven to move to the next sow pen. Among them, the fine-tuning parameters of the biomimetic arousal mechanism include the control parameters for generating the modulation control sequence, the audio modulation parameters, and the release frequency of the odor pulse.

9. The intelligent bionic arousal and arousal detection robot system according to claim 8, characterized in that, In the multimodal cross-attention fusion module, multidimensional response features are fused to form fused features, including: Flatten the visual features into a query tensor The thermal infrared features and audio features are concatenated into a key. Sum Tensors, through calculation The feature weights of thermal infrared features and audio features are determined, and then the visual features, thermal infrared features and audio features are spliced ​​together to obtain the fused features.

10. A working method of an intelligent biomimetic arousal-inducing and arousal-detecting robot system, characterized in that, include: S100, Global Initialization: Wake up the system and perform sensor calibration, and perform a panoramic scan of the pigsty during the movement of the mobile adapter chassis; S200, Target Locking: Determine the face and ear mask of the target sow, continuously calculate the real-time distance from the scent nozzle of the estrus detection robot to the sow's snout, and activate follow-up tracking; S300, Bionic stimulus: Responds to the target mask exceeding a set consecutive frames when the target sow detection box confidence is greater than a set threshold. A four-dimensional hyperchaotic system is activated to generate a modulation control sequence and dynamically modulates pre-recorded pig audio samples, thereby transmitting anti-habitation directional estrus-inducing audio. The real-time distance is input into the inverse model trained based on the distance compensation relationship, and the corresponding duty cycle PWM control signal is output to control the solenoid valve of the odor nozzle to release odor irritants and perform micro-distance approach-backward probing action. S400, Feature Extraction: During the observation window of continuous excitation, visual features, thermal infrared features, and acoustic features are extracted and fused and compensated to generate fused features; S500, Estrus Diagnosis: After the observation window period ends, the fused features and time indicators are input into the LoRA large model. Estrus diagnosis is performed based on the duration of static standing with both ears, the duration of active contact with the estrus-inducing robot, and the consistency of multimodal features, and a diagnostic report is generated. In response to the diagnosis that the sow is not in estrus, the control parameters of the modulation control sequence for generating the variable frequency sound wave, the audio modulation parameters, and the release frequency of the odor pulse are finely adjusted, and the process returns to S300.

Images