A multi-source information fusion intelligent identification method and system for sheep health monitoring

Abstract

The present application relates to the technical field of animal-related pattern recognition in image / video data, and particularly relates to a multi-source information fusion intelligent recognition method and system for sheep health monitoring. The method first obtains a pair of visual data through synchronous verification and pairing of visible light and infrared acquisition units, and screens high-quality data through preprocessing and quality verification. Then, features are extracted, redundant and weakly distinguishable features are removed, and the fusion weight is dynamically adjusted in combination with the feature consistency coefficient to obtain fused features. Finally, the prediction result is obtained through hierarchical decision fusion, the sheep health and emotional state are determined based on the confidence, and the state duration is combined for grading early warning. The method ensures efficient fusion of multi-source information, improves recognition accuracy, ensures timely and accurate early warning, adapts to the needs of fine monitoring of large-scale breeding, reduces breeding risks, and improves efficiency.

Description

Technical Field

[0001] This invention relates to the field of animal-related pattern recognition technology in image / video data, and more specifically, to a multi-source information fusion intelligent recognition method and system for sheep health monitoring. Background Technology

[0002] With the rapid development of large-scale and intensive sheep farming, sheep health and emotional state are directly related to farming efficiency and product quality. Traditional manual monitoring methods are limited by high labor costs, strong subjectivity, and slow response, making them difficult to adapt to the monitoring needs of large-scale farming scenarios. Intelligent monitoring technology is urgently needed to break through the bottleneck.

[0003] Existing sheep monitoring technologies mostly rely on the collection and analysis of data from a single visual modality. Although some attempts have been made to integrate information from multiple sources, there is a lack of precise control over the synchronization of the collection units. Furthermore, redundant and weakly distinguishable features are not effectively removed during the feature extraction stage, and the fusion weights are mostly fixed settings, which cannot adapt to the dynamic changes in the breeding environment and the prediction differences of different feature modalities.

[0004] The current core technical problem is that existing solutions cannot achieve efficient and accurate fusion of visible light and infrared multi-source visual information, resulting in insufficient accuracy in recognizing sheep health and emotional state, and the inability to output graded early warning information in a timely and accurate manner, which makes it difficult to meet the refined monitoring needs of large-scale farming. Summary of the Invention

[0005] In view of this, the present invention proposes a multi-source information fusion intelligent identification method and system for sheep health monitoring, which aims to solve the problem that the existing technology is difficult to achieve efficient and accurate fusion of visible light and infrared multi-source visual information, resulting in insufficient accuracy in identifying sheep health and emotional state, inability to output graded early warning information in a timely and accurate manner, and difficulty in meeting the refined monitoring needs of large-scale breeding.

[0006] In a first aspect, the present invention provides a multi-source information fusion intelligent identification method for sheep health monitoring, comprising the following steps: A visible light acquisition unit and an infrared acquisition unit are deployed to collect visual data of sheep in the target sheep breeding area. The visible light acquisition unit and the infrared acquisition unit are synchronized. Based on the verification result, it is determined whether to perform unit synchronization calibration. The visible light image and the infrared image that pass the verification are paired to form a visual data pair. The visual data pairs are preprocessed, and a visible light image sharpness threshold and an infrared image temperature validity threshold are preset. The quality of the preprocessed visual data pairs is checked according to the preset thresholds, and it is determined whether to re-acquire data based on the check results. Based on the preprocessed visual data pairs, visible light feature sets and infrared feature sets are extracted respectively. The correlation degree of each feature dimension in the two feature sets is calculated to determine whether the feature pairs are redundant. The redundancy feature pairs are filtered by calculating the discrimination degree, and the discrimination degree of each feature after filtering is calculated. Weak discrimination features are removed according to the preset weak discrimination threshold to obtain the multimodal effective feature set. The effective features of visible light and infrared light are normalized and initial fusion weights are set to obtain initial fusion features. The consistency of the prediction results of the two feature modes is judged by the feature consistency coefficient. The fusion weights are dynamically adjusted and the fusion features are updated according to the range of the feature consistency coefficient. The final prediction result is obtained by combining hierarchical decision fusion. The health and emotional state of sheep are determined according to the preset confidence threshold. The warning level is divided according to the duration of the state and the corresponding warning information is output.

[0007] In some embodiments, the deployment of visible light acquisition units and infrared acquisition units for collecting visual data of sheep in a target sheep breeding area, performing synchronization verification on the visible light acquisition units and infrared acquisition units, determining whether to perform unit synchronization calibration based on the verification results, and pairing the verified visible light images with infrared images to form visual data pairs includes: Based on the sheep's walking speed, a time difference threshold for acquisition is determined, and it is determined whether the acquisition time difference between the visible light image and the infrared image is less than or equal to the acquisition time difference threshold. When the acquisition time difference between the visible light image and the infrared image is greater than the acquisition time difference threshold, a synchronization check is performed, and a synchronization check is performed again after calibration. If the acquisition time difference between the visible light image and the infrared image after secondary calibration is greater than the acquisition time difference threshold, a first warning signal will be issued. If the acquisition time difference between the visible light image and the infrared image after secondary calibration is less than or equal to the acquisition time difference threshold, then the data pair is determined to be valid.

[0008] In some embodiments, the preprocessing of the visual data pairs includes: Gaussian filtering is applied to the visible light image to remove environmental noise, and adaptive histogram equalization is then used to enhance the contrast between fur texture and body contour. The infrared image is filtered by median filtering to remove thermal noise, and then the body temperature distribution is transformed into a visual feature through pseudo-color mapping. At the same time, the body temperature data matrix corresponding to the infrared image is extracted. The filtering parameters of the visible light image preprocessing are dynamically adjusted according to the ambient brightness of the target sheep breeding area, and the noise removal intensity of the infrared image preprocessing is dynamically adjusted according to the ambient temperature of the target sheep breeding area.

[0009] In some embodiments, the preprocessing of the visual data pairs, and the setting of a visible light image sharpness threshold and an infrared image temperature validity threshold, followed by quality verification of the preprocessed visual data pairs based on the preset thresholds, and determining whether to re-acquire data based on the verification result, includes: The sharpness of the visible light image is calculated using the gray-scale variance method, and a threshold for the sharpness of the visible light image is preset. The temperature validity of the infrared image is calculated based on the percentage of effective temperature pixels, and a preset threshold for the temperature validity of the infrared image is established. When the visible light image sharpness in the preprocessed visual data pair is greater than or equal to the visible light image sharpness threshold, and the infrared image temperature validity is greater than or equal to the infrared image temperature validity threshold, it is determined to be a high-quality image; When the visible light image sharpness in the preprocessed visual data pair is greater than or equal to the visible light image sharpness threshold, and the infrared image temperature validity is less than the infrared image temperature validity threshold, the infrared image is reacquired. When the visible light image sharpness in the preprocessed visual data pair is less than the visible light image sharpness threshold, and the infrared image temperature validity is greater than or equal to the infrared image temperature validity threshold, the visible light image is re-acquired; When the visible light image sharpness in the preprocessed visual data pair is less than the visible light image sharpness threshold, and the infrared image temperature validity is less than the infrared image temperature validity threshold, the visible light image and the infrared image are reacquired.

[0010] In some embodiments, when extracting visible light feature sets and infrared feature sets based on the preprocessed visual data pairs, determining whether a feature pair is redundant by calculating the correlation degree of each feature dimension in the two feature sets, calculating the discrimination of redundant feature pairs for filtering, calculating the discrimination of each feature after filtering, and removing weakly discriminative features according to a preset weak discrimination threshold to obtain a multimodal effective feature set, the process includes: When the correlation between two feature dimensions is greater than a preset redundancy threshold, they are determined to be a redundant feature pair. When the correlation between two feature dimensions is less than or equal to the redundancy threshold, they are determined to be non-redundant feature pairs and are directly retained. The discrimination index is obtained by calculating the mean difference and variance percentage of corresponding characteristics between healthy sheep and abnormal sheep.

[0011] In some embodiments, when extracting visible light feature sets and infrared feature sets based on the preprocessed visual data pairs, determining whether a feature pair is redundant by calculating the correlation degree of each feature dimension in the two feature sets, calculating the discrimination of redundant feature pairs for filtering, calculating the discrimination of each feature after filtering, and removing weakly discriminative features according to a preset weak discrimination threshold to obtain a multimodal effective feature set, the process includes: A weak discrimination threshold is preset, the discrimination of each feature after filtering is calculated, and it is determined whether the discrimination of each feature is greater than or equal to the weak discrimination threshold. When the feature discrimination is less than the weak discrimination threshold, it is determined to be a weak discrimination feature and is removed. When the feature discrimination is greater than or equal to the weak discrimination threshold, it is determined to be a valid feature and retained in the multimodal valid feature set; After the elimination is completed, it is determined whether the dimension of the multimodal effective feature set is greater than or equal to the preset minimum dimension requirement. If it is not met, the features are extracted again.

[0012] In some embodiments, the process of normalizing visible light effective features and infrared effective features and setting initial fusion weights to obtain initial fusion features, introducing a feature consistency coefficient to judge the consistency of prediction results between the two feature modes, dynamically adjusting the fusion weights and updating the fusion features according to the range of the feature consistency coefficient, combining hierarchical decision fusion to obtain the final prediction result, and then determining the sheep's health and emotional state according to a preset confidence threshold, dividing the warning level according to the state duration and outputting the corresponding warning information includes: A high consistency threshold and a medium consistency threshold are preset. It is determined whether the feature consistency coefficient is greater than or equal to the high consistency threshold. When the feature consistency coefficient is greater than or equal to the high consistency threshold, the initial fusion weight is maintained. When the feature consistency coefficient is less than the high consistency threshold and greater than or equal to the medium consistency threshold, the historical prediction accuracy of the two feature modalities is calculated, and the fusion weight is adjusted according to the accuracy ratio. When the feature consistency coefficient is less than the consistency threshold, the discriminative power of the two feature modes is determined, and the weight corresponding to the feature mode with higher discriminative power is greater than the weight corresponding to the feature mode with lower discriminative power. When two feature modes have the same discriminative power, the weights corresponding to the two feature modes are the same.

[0013] In some embodiments, the process of normalizing visible light effective features and infrared effective features and setting initial fusion weights to obtain initial fusion features, judging the consistency of prediction results between two feature modes by feature consistency coefficient, dynamically adjusting fusion weights and updating fusion features according to the range of feature consistency coefficient, obtaining the final prediction result by combining hierarchical decision fusion, determining the sheep's health and emotional state according to a preset confidence threshold, classifying warning levels and outputting corresponding warning information based on the duration of the state, includes: The updated fused features are input into the machine learning model to obtain preliminary prediction results. At the same time, the visible light effective features and infrared effective features are input into the same machine learning model to obtain two single-modal prediction results. The weight of the preliminary prediction result is greater than the weight of the two single-modality prediction results; When the deviation of the prediction results for a preset number of consecutive predictions is less than a preset deviation threshold, it is determined to be the final prediction result; If the deviation of the prediction results for a preset number of consecutive predictions is greater than or equal to the deviation threshold, then the fusion calculation is performed again.

[0014] In some embodiments, the process of normalizing visible light effective features and infrared effective features and setting initial fusion weights to obtain initial fusion features, judging the consistency of prediction results between two feature modes by feature consistency coefficient, dynamically adjusting fusion weights and updating fusion features according to the range of feature consistency coefficient, obtaining the final prediction result by combining hierarchical decision fusion, determining the sheep's health and emotional state according to a preset confidence threshold, classifying warning levels and outputting corresponding warning information based on the duration of the state, includes: A high confidence threshold and a suspected confidence threshold are preset. It is determined whether the confidence of the final prediction result is greater than or equal to the high confidence threshold. When the confidence of the final prediction result is greater than or equal to the high confidence threshold, the corresponding sheep health and emotional state labels are output. If the confidence level of the final prediction result is greater than or equal to the suspected confidence level threshold and less than the high confidence level threshold, then a second data collection is performed. When the confidence level of the final prediction result is greater than or equal to the suspected confidence threshold and less than the high confidence threshold after the second data collection, a second warning signal is issued. If the confidence level of the final prediction result is less than the suspected confidence level threshold, the data is determined to be abnormal and the data is collected again. Preset short-term duration thresholds and long-term duration thresholds. Based on the type of abnormal state and whether the duration is greater than or equal to the corresponding threshold, the system determines the level of warning as Level 1, Level 2, or Level 3, respectively.

[0015] Secondly, the present invention provides a multi-source information fusion intelligent identification system for sheep health monitoring, comprising: The acquisition module is configured to deploy a visible light acquisition unit and an infrared acquisition unit for acquiring visual data of sheep in the target sheep breeding area. The visible light acquisition unit and the infrared acquisition unit are synchronized. Based on the verification result, it is determined whether to perform unit synchronization calibration. The visible light image and the infrared image that pass the verification are paired to form a visual data pair. The verification module is configured to preprocess the visual data pairs, preset a visible light image sharpness threshold and an infrared image temperature validity threshold, perform quality verification on the preprocessed visual data pairs according to the preset thresholds, and determine whether to re-acquire data based on the verification results. The filtering module is configured to extract visible light feature sets and infrared feature sets based on the preprocessed visual data pairs, determine whether the feature pairs are redundant by calculating the correlation degree of each feature dimension in the two feature sets, calculate the discrimination degree of redundant feature pairs and filter them, calculate the discrimination degree of each feature after filtering, and remove weakly discriminative features according to a preset weak discrimination threshold to obtain a multimodal effective feature set. The early warning module is configured to normalize the effective features of visible light and infrared light and set initial fusion weights to obtain initial fusion features. It judges the consistency of the prediction results of the two feature modes by the feature consistency coefficient, dynamically adjusts the fusion weights and updates the fusion features according to the range of the feature consistency coefficient, and obtains the final prediction result by combining hierarchical decision fusion. Then, it determines the health and emotional state of the sheep according to the preset confidence threshold, divides the early warning level according to the state duration and outputs the corresponding early warning information.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: Through the synchronization verification and pairing design of visible light and infrared acquisition units, the spatiotemporal consistency of multi-source visual data is ensured, solving the data mismatch problem caused by insufficient synchronization control in existing solutions, and laying a high-quality data foundation for subsequent fusion analysis. Preprocessing and quality verification mechanisms effectively eliminate low-quality data, improving the reliability of input data and avoiding interference from invalid data on recognition results. Redundant feature removal and weakly discriminative feature screening simplify feature dimensions and strengthen effective feature representation. Combined with dynamic adjustment of fusion weights using feature consistency coefficients, the invention overcomes the poor adaptability of fixed weights, achieving efficient and accurate fusion of multimodal information and significantly improving the accuracy of sheep health and emotional state recognition. Hierarchical decision fusion and graded early warning design based on confidence level and duration ensure timely and accurate early warning information, adapting to the refined monitoring needs of large-scale farming, reducing farming risks, and improving farming efficiency.

[0017] The above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0018] Other features and aspects of this disclosure will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a multi-source information fusion intelligent identification method for sheep health monitoring provided in an embodiment of the present invention; Figure 2 This is a flowchart of a multi-source information fusion intelligent identification method for sheep health monitoring provided in an embodiment of the present invention; Figure 3 This is a functional block diagram of a multi-source information fusion intelligent identification system for sheep health monitoring provided in an embodiment of the present invention. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] See Figure 1-2 As shown in the first embodiment, a multi-source information fusion intelligent identification method for sheep health monitoring according to an embodiment of this application is characterized by including the following steps: A visible light acquisition unit and an infrared acquisition unit are deployed to collect visual data of sheep in the target sheep breeding area. The visible light acquisition unit and the infrared acquisition unit are synchronized. Based on the verification result, it is determined whether to perform unit synchronization calibration. The visible light image and the infrared image that pass the verification are paired to form a visual data pair. The visual data pairs are preprocessed, and a visible light image sharpness threshold and an infrared image temperature validity threshold are preset. The quality of the preprocessed visual data pairs is checked according to the preset thresholds, and it is determined whether to re-acquire data based on the check results. Based on the preprocessed visual data pairs, visible light feature sets and infrared feature sets are extracted respectively. The correlation degree of each feature dimension in the two feature sets is calculated to determine whether the feature pairs are redundant. The redundancy feature pairs are filtered by calculating the discrimination degree, and the discrimination degree of each feature after filtering is calculated. Weak discrimination features are removed according to the preset weak discrimination threshold to obtain the multimodal effective feature set. The effective features of visible light and infrared light are normalized and initial fusion weights are set to obtain initial fusion features. The consistency of the prediction results of the two feature modes is judged by the feature consistency coefficient. The fusion weights are dynamically adjusted and the fusion features are updated according to the range of the feature consistency coefficient. The final prediction result is obtained by combining hierarchical decision fusion. The health and emotional state of sheep are determined according to the preset confidence threshold. The warning level is divided according to the duration of the state and the corresponding warning information is output.

[0023] It should be understood that the core technical means of this solution consists of a data acquisition subsystem, a preprocessing and verification subsystem, a feature processing subsystem, and a fusion decision and early warning subsystem. The data acquisition subsystem includes a visible light acquisition unit (high-definition industrial camera, resolution 1920×1080, frame rate 25fps) and an infrared acquisition unit (uncooled infrared thermal imager, temperature measurement range -20℃~150℃, pixel 320×240). Both integrate a timestamp synchronization module. Synchronization verification is achieved by comparing the timestamp difference between the data acquired by the two units, and synchronization calibration is completed by adjusting the timing parameters of the acquisition trigger signal. In the preprocessing and verification subsystem, the preprocessing module employs Gaussian filtering (with dynamically adjustable standard deviation) and adaptive histogram equalization for visible light images, and median filtering and pseudo-color mapping (using the Jet color mapping table) for infrared images, while simultaneously extracting the body temperature data matrix. The quality verification module calculates the sharpness of visible light images using the grayscale variance method and calculates the temperature validity of infrared images by the percentage of effective temperature pixels (the proportion of pixels within the normal body temperature range of sheep, 38.5-39.5℃, out of the total number of pixels). Preset thresholds are determined based on statistical analysis of numerous livestock farming samples. The feature processing subsystem uses the SIFT algorithm to extract texture and contour features from visible light images, employs histogram statistics to extract temperature distribution features from infrared images, calculates feature correlation using the Pearson correlation coefficient to determine redundancy, and calculates the discriminative power by using the ratio of "difference between the mean features of healthy and abnormal sheep / total variance of the two types of features," eliminating weakly discriminative features. The fusion decision-making and early warning subsystem first uses min-max normalization to process effective features. The initial fusion weights are set based on expert experience (0.4 for visible light and 0.6 for infrared). The feature consistency coefficient is obtained by calculating the cosine similarity between the two modal prediction results. Dynamic weight adjustment triggers different rules based on the range of the consistency coefficient. The hierarchical decision fusion adopts a weight allocation strategy of "fusion feature prediction results as the main factor and single modal prediction results as the auxiliary factor". The confidence level is the probability value of the model prediction result, and the early warning level is set according to the correspondence between the state duration and the anomaly type. The subsystems interact with each other through industrial Ethernet, and the core processing module uses a GPU (NVIDIA Jetson AGXXavier) to ensure real-time performance.

[0024] Taking a large-scale sheep farm (500 sheep and 1000㎡ of breeding area) as an example, four sets of visible light-infrared acquisition units were evenly deployed in the breeding area. The statistical value of sheep walking speed was set to 0.5m / s. The acquisition time difference threshold was derived to be 0.2s, the visible light image clarity threshold was 80, the infrared image temperature validity threshold was 70%, the redundancy feature threshold was 0.7, the weak discrimination threshold was 0.3, the feature consistency coefficient high threshold was 0.9, medium threshold was 0.7, high confidence threshold was 0.9, suspected confidence threshold was 0.7, short duration threshold was 2h, and long duration threshold was 8h. During system operation, the acquisition unit collects data every 10 seconds. Synchronization verification revealed a 0.3-second time difference in a certain data set, triggering synchronization calibration which reduced the time difference to 0.15 seconds, forming a valid data pair. In the preprocessing stage, the Gaussian filter standard deviation was adjusted to 1.2 for cloudy conditions (ambient brightness 100 lux), and the median filter intensity for infrared images was enhanced for an ambient temperature of 28℃. After processing, the visible light image texture contrast improved by 30%, and the infrared image thermal noise decreased by 25%. Quality verification revealed a visible light image with a sharpness of 75 (below the threshold), triggering re-acquisition. Feature extraction yielded 32 dimensions of visible light features and 24 dimensions of infrared features, resulting in the calculation of 12... For features with a correlation greater than 0.7, 8 pairs of high-discrimination features were retained through discrimination screening, and 6 weak-discrimination features with a discrimination of less than 0.3 were removed, resulting in an effective feature set of 40 dimensions. After normalization, the initial fusion weights were 0.4-0.6, and the feature consistency coefficient was calculated to be 0.8 (between the medium and high thresholds). Based on the historical prediction accuracy (85% for visible light and 90% for infrared), the weights were adjusted to 0.36-0.64. After hierarchical decision fusion, the prediction result was obtained with a confidence level of 0.92 (higher than the high confidence threshold), which was judged as a "mild stress state". No warning was triggered for 1.5 hours (not reaching the short-term threshold), and a level 1 warning was triggered after 2.5 hours.The technical solutions offer significant advantages: First, they ensure the quality of multi-source data. Synchronization verification and calibration reduce the spatiotemporal consistency error of data to within 0.05 seconds, and after quality verification and screening, the proportion of high-quality data increases to over 92%, effectively avoiding interference from invalid data in subsequent processing. Second, they optimize feature representation capabilities. Redundant and weakly discriminative features are eliminated, reducing feature dimensions by 40% and improving feature computation efficiency by 50%. Simultaneously, the representation capabilities of effective features are enhanced, laying the foundation for accurate fusion. Third, they improve recognition and early warning accuracy. Dynamic weight adjustment enables the recognition accuracy of multimodal fusion to reach over 96%, significantly higher than single-modal recognition. The accuracy rate is improved by 15-20%, and the hierarchical decision-making and graded early warning mechanism shortens the early warning response time to less than 30 seconds, with a false negative rate of less than 2% and a false positive rate of less than 3%. Fourth, it is adapted to the needs of large-scale farming. The system can simultaneously cover multiple areas and monitor multiple sheep without human intervention, reducing labor costs by more than 60%. Timely early warning can reduce the mortality rate of sheep due to health problems by more than 5%, significantly improving farming efficiency. Fifth, it has environmental adaptability. The dynamically adjusted preprocessing parameters ensure that the system's recognition stability remains above 90% under different environments such as sunny days, cloudy days, high temperatures, and low temperatures, avoiding performance fluctuations caused by environmental changes.

[0025] In some specific embodiments, the deployment of visible light acquisition units and infrared acquisition units for collecting visual data of sheep in the target sheep breeding area, the synchronization verification of the visible light acquisition units and infrared acquisition units, the determination of whether to perform unit synchronization calibration based on the verification result, and the pairing of the verified visible light image and infrared image to form a visual data pair include: Based on the sheep's walking speed, a time difference threshold for acquisition is determined, and it is determined whether the acquisition time difference between the visible light image and the infrared image is less than or equal to the acquisition time difference threshold. When the acquisition time difference between the visible light image and the infrared image is greater than the acquisition time difference threshold, a synchronization check is performed, and a synchronization check is performed again after calibration. If the acquisition time difference between the visible light image and the infrared image after secondary calibration is greater than the acquisition time difference threshold, a first warning signal will be issued. If the acquisition time difference between the visible light image and the infrared image after secondary calibration is less than or equal to the acquisition time difference threshold, then the data pair is determined to be valid.

[0026] It should be understood that the core is to determine the acquisition time difference threshold based on the sheep's walking speed, and to determine whether the acquisition time difference between visible light and infrared images meets the requirements through two synchronization checks. If it does not meet the requirements, synchronization calibration is performed. If it still does not meet the requirements after the second calibration, the first warning signal is issued. If it meets the requirements, the data pair is determined to be valid, ensuring that the formed visual data pair has spatiotemporal consistency and providing a reliable data foundation for subsequent multi-source information fusion.

[0027] The key technical approach of this solution is a closed-loop mechanism of "sheep walking speed statistics - time difference threshold calculation - two synchronization verifications - synchronization calibration - early warning triggering". Sheep walking speed is collected using speed monitoring sensors (LiDAR sensors, measurement accuracy ±0.01m / s) deployed in the breeding area, collecting 1000 sets of sheep movement data under different scenarios (feeding, walking, resting). The average walking speed v is obtained through statistical analysis. Based on this, the time difference threshold t is calculated using the formula t=d / v, where d is the installation distance between the visible light and infrared acquisition units (fixed at 0.1m), ensuring that the displacement of the same sheep within the acquisition range of the two units does not affect the accuracy of data pairing. Synchronization verification involves reading the timestamps of the data collected by the two acquisition units (accurate to the millisecond level), calculating the timestamp difference Δt, and comparing Δt with the threshold t. Synchronous calibration employs a "hardware trigger + software fine-tuning" approach. At the hardware level, a synchronization trigger module (with an adjustable signal delay range of 0-1s and an adjustment precision of 0.01s) uniformly controls the acquisition trigger signals of both units. At the software level, a timestamp compensation algorithm fine-tunes the timestamps of the acquired data to ensure the time difference meets requirements after calibration. The first warning signal is "acquisition unit synchronization abnormality," issued through both a local audible and visual alarm (alarm sound level ≥85dB, alarm light flashing red) and a pop-up notification on the remote monitoring platform. Simultaneously, the abnormal time, unit number, and other information are recorded for easy troubleshooting by maintenance personnel.

[0028] Taking a sheep farm as an example, the average walking speed of the sheep is v=0.5m / s collected by a lidar sensor. The installation spacing of the acquisition units is d=0.1m. Substituting these values ​​into the formula, the acquisition time difference threshold t=0.2s is calculated. During system operation, a set of acquisition units acquires a visible light image with a timestamp of 1699999999.123s, corresponding to an infrared image timestamp of 1699999999.456s. The calculated time difference Δt=0.333s>0.2s triggers a synchronization verification anomaly, initiating synchronization calibration: at the hardware level, the signal delay of the synchronization trigger module is adjusted, advancing the trigger signal of the infrared acquisition unit by 0.2s; at the software level, the infrared image timestamp is compensated and corrected to 1699999999.256s. Δt=0.133s≤0.2s is calculated again, indicating that the calibration is valid and the data pairing is successful. If, after the first calibration of another set of acquisition units, Δt = 0.22s > 0.2s, a second calibration is initiated, and the trigger signal delay is adjusted to 0.25s. If, after correction, Δt = 0.21s is still > 0.2s, then a first warning signal is issued to remind maintenance personnel to check the hardware connection and synchronization module working status of the acquisition units. This technical solution offers several key advantages: First, it accurately determines the time difference threshold. Based on statistical calculations of sheep's actual walking speed, the threshold better reflects the realities of the farming scenario, avoiding synchronization errors caused by fixed thresholds and improving synchronization accuracy by 25% compared to fixed threshold solutions. Second, it ensures data spatiotemporal consistency. The dual synchronization verification and precise calibration mechanism increases the proportion of data meeting time difference requirements to over 98%, effectively preventing data mismatches for the same sheep due to excessive time differences and reducing error sources in subsequent fusion analysis. Third, it promptly detects synchronization anomalies. The dual alert method of the first warning signal reduces the response time for maintenance personnel to synchronization anomalies to within 5 minutes, reducing the accumulation of invalid data caused by synchronization anomalies and ensuring continuous and stable system operation. Fourth, it improves data pairing efficiency. The automated synchronization verification and calibration process requires no manual intervention, reducing data pairing time to within 0.5 seconds, adapting to the real-time monitoring needs of large-scale farming scenarios. Fifth, it reduces the impact of hardware failures. The synchronization anomaly warning system proactively identifies and addresses acquisition unit failures, reducing system downtime due to hardware issues and improving overall system availability.

[0029] In some specific embodiments, the preprocessing of the visual data pairs includes: Gaussian filtering is applied to the visible light image to remove environmental noise, and adaptive histogram equalization is then used to enhance the contrast between fur texture and body contour. The infrared image is filtered by median filtering to remove thermal noise, and then the body temperature distribution is transformed into a visual feature through pseudo-color mapping. At the same time, the body temperature data matrix corresponding to the infrared image is extracted. The filtering parameters of the visible light image preprocessing are dynamically adjusted according to the ambient brightness of the target sheep breeding area, and the noise removal intensity of the infrared image preprocessing is dynamically adjusted according to the ambient temperature of the target sheep breeding area.

[0030] It should be understood that Gaussian filtering and adaptive histogram equalization are used to enhance visible light images, while median filtering and pseudo-color mapping visualization are used to extract body temperature data matrices for infrared images. At the same time, the visible light preprocessing filtering parameters are dynamically adjusted according to the ambient brightness of the breeding area, and the infrared preprocessing noise removal intensity is dynamically adjusted according to the ambient temperature to ensure that the preprocessed data has good feature recognition and is suitable for subsequent feature extraction needs.

[0031] The core technology of this solution is the "modal-specific preprocessing + dynamic adaptation of environmental parameters" mechanism. In visible light image preprocessing, Gaussian filtering removes Gaussian noise caused by ambient light interference by constructing a Gaussian kernel function (kernel size adjustable from 3×3 / 5×5 / 7×7) and performing convolution operations on the image. The filtering parameters (kernel size, standard deviation σ) are negatively correlated with the ambient brightness, which is acquired in real time by a light sensor (measurement range 0-10000 lux, accuracy ±1 lux). Adaptive histogram equalization divides the image into multiple sub-regions (sub-region size 8×8) and performs histogram equalization on each sub-region separately, avoiding the loss of local details caused by global equalization and enhancing the contrast of fur texture (such as wool thickness and color distribution) and body contour (such as body shape and limb posture). In infrared image preprocessing, median filtering removes salt-and-pepper noise caused by thermal radiation by replacing the original pixel value with the median value within the pixel's neighborhood. The noise removal intensity is achieved by adjusting the filter kernel size (3×3 / 5×5), which is positively correlated with the ambient temperature. The ambient temperature is acquired in real time by a temperature sensor (measurement range -30℃~50℃, accuracy ±0.1℃). Pseudo-color mapping uses the Jet color mapping table to map different temperature values ​​to different colors (low temperature corresponds to blue, high temperature corresponds to red), making body temperature distribution visible. The body temperature data matrix is ​​formed by reading the raw temperature data output from the infrared thermal imager and arranging it according to the image pixel coordinates to form a two-dimensional matrix (320×240). The matrix element value is the actual temperature of the corresponding pixel. The mapping relationship between environmental parameters and preprocessing parameters is obtained through fitting a large amount of experimental data and stored in the system parameter configuration library. The preprocessing module reads the environmental sensor data in real time and automatically matches the corresponding preprocessing parameters.

[0032] Taking different environmental scenarios of a sheep farm as an example, Scenario 1: Sunny noon, ambient brightness 8000 lux. After the light sensor collects data, the system automatically matches the visible light preprocessing Gaussian filter parameters (kernel size 3×3, σ=0.8). After adaptive histogram equalization, the grayscale contrast of sheep wool texture is improved by 40%, and the edge clarity of the body contour is improved by 35%. The ambient temperature is 32℃. The temperature sensor data triggers the infrared preprocessing median filter kernel size of 3×3. After processing, the number of salt and pepper noise points in the infrared image is reduced by 60%. After pseudo-color mapping, the temperature difference of different parts of the sheep's body can be clearly distinguished. The body temperature data matrix accurately records the temperature of each part (e.g., ears 39.2℃, back 38.8℃). Scenario 2: On a cloudy morning with an ambient brightness of 500 lux, the system matches the visible light filtering parameters (kernel size 5×5, σ=1.5). After filtering, the Gaussian noise removal rate of the image reaches 85%. After adaptive histogram equalization, the outline of the sheep can still be clearly identified even in low-brightness environments. The ambient temperature is 15℃. The infrared filter kernel size is matched to 3×3. After processing, the thermal noise of the infrared image is reduced by 55%, and the error of the body temperature data matrix is ​​controlled within ±0.1℃. The technical solutions offer significant advantages: First, they improve image preprocessing quality, achieving a noise reduction rate of over 80% for visible light images and a contrast improvement of 30-40%, while achieving a noise reduction rate of over 75% for infrared images. This results in excellent body temperature visualization, providing high-quality data for subsequent feature extraction. Second, they enhance environmental adaptability. Dynamically adjusted preprocessing parameters maintain a preprocessing effect stability of over 90% under varying brightness (0-10000 lux) and temperature (-30℃~50℃) conditions, preventing preprocessing failure under extreme environments with fixed parameters. Third, they strengthen feature recognition. By enhancing the contrast between coat texture and body contour and clarifying body temperature distribution, the accuracy of subsequent feature extraction is improved by over 20%, reducing recognition errors caused by feature blurring. Fourth, they ensure the accuracy of body temperature data. The error of the body temperature data matrix after infrared image preprocessing is controlled within ±0.1℃, providing reliable data support for temperature-based health status assessment. Fifth, they improve the automation level of preprocessing, eliminating the need for manual parameter adjustments, adapting to the all-weather monitoring needs of large-scale farming scenarios, and reducing manual maintenance costs.

[0033] In some specific embodiments, the preprocessing of the visual data pairs, including setting a visible light image sharpness threshold and an infrared image temperature validity threshold, performing quality checks on the preprocessed visual data pairs based on the preset thresholds, and determining whether to re-acquire data based on the check results, includes: The sharpness of the visible light image is calculated using the gray-scale variance method, and a threshold for the sharpness of the visible light image is preset. The temperature validity of the infrared image is calculated based on the percentage of effective temperature pixels, and a preset threshold for the temperature validity of the infrared image is established. When the visible light image sharpness in the preprocessed visual data pair is greater than or equal to the visible light image sharpness threshold, and the infrared image temperature validity is greater than or equal to the infrared image temperature validity threshold, it is determined to be a high-quality image; When the visible light image sharpness in the preprocessed visual data pair is greater than or equal to the visible light image sharpness threshold, and the infrared image temperature validity is less than the infrared image temperature validity threshold, the infrared image is reacquired. When the visible light image sharpness in the preprocessed visual data pair is less than the visible light image sharpness threshold, and the infrared image temperature validity is greater than or equal to the infrared image temperature validity threshold, the visible light image is re-acquired; When the visible light image sharpness in the preprocessed visual data pair is less than the visible light image sharpness threshold, and the infrared image temperature validity is less than the infrared image temperature validity threshold, the visible light image and the infrared image are reacquired.

[0034] It should be understood that the sharpness of visible light images is calculated by the gray-scale variance method, and the temperature validity of infrared images is calculated by the effective temperature pixel ratio. Corresponding sharpness thresholds and temperature validity thresholds are preset. Based on the verification results of the preprocessed data, infrared image re-acquisition, visible light image re-acquisition, dual image re-acquisition, or high-quality image determination are triggered respectively. This ensures that the data input to the subsequent feature extraction stage is high-quality data, thereby improving the accuracy of feature extraction and recognition.

[0035] The core technology of this solution is a "dual-dimensional quality assessment + graded re-acquisition triggering" mechanism. Visible light image sharpness is calculated using the gray-level variance method, where the gray-level variance σ... 2 The calculation formula is σ 2 =Σ(x ij -μ) 2 / (M×N), where x ijLet μ be the grayscale value of the pixel at coordinate (i,j) in the image, μ be the average grayscale value of the image, and M×N be the image resolution (1920×1080). The larger the grayscale variance, the clearer the image. The clarity threshold is determined by statistically analyzing the grayscale variance of 1000 high-quality visible light images and taking the minimum value as the threshold (set to 80 in this embodiment). The temperature validity of the infrared image is calculated by the percentage of effective temperature pixels. Effective temperature pixels are defined as pixels whose temperature falls within the normal body temperature range of sheep (38.5-39.5℃). The percentage is calculated as (number of effective pixels / total number of pixels) × 100%. The temperature validity threshold is determined by statistically analyzing the percentage of effective temperature pixels in 1000 high-quality infrared images and taking the minimum value as the threshold (set to 70% in this embodiment). The quality verification module receives preprocessed visual data pairs in real time, calculates sharpness and temperature validity separately, and compares them with preset thresholds to generate four verification results: both parameters meet the standard (high-quality image), visible light meets the standard but infrared does not (re-acquire infrared), visible light does not meet the standard but infrared meets the standard (re-acquire visible light), and both parameters do not meet the standard (re-acquire dual images). After re-acquisition is triggered, the system sends a re-acquisition command to the corresponding acquisition unit. The acquisition unit completes re-acquisition within 1 second. The maximum number of re-acquisitions is 3. If the standard is still not met after 3 attempts, a "data acquisition anomaly" warning is issued.

[0036] Taking the visual data quality verification process of a sheep farm as an example, Case 1: After preprocessing, the grayscale variance of the visible light image is 95 (≥ threshold 80), and the effective temperature pixel ratio of the infrared image is 78% (≥ threshold 70%), which is judged as a high-quality image and directly input into the feature extraction stage. Case 2: The grayscale variance of the visible light image is 92 (≥ 80), and the effective temperature pixel ratio of the infrared image is 65% (< 70%). The system sends a re-acquisition command to the infrared acquisition unit. After re-acquisition, the effective temperature pixel ratio is calculated to be 76% (≥ 70%), and the data is qualified. Case 3: The grayscale variance of the visible light image is 72 (< 80), and the effective temperature pixel ratio of the infrared image is 75% (≥ 70%). Re-acquisition of the visible light image is triggered. The grayscale variance of the first re-acquisition is 76 (< 80), and the grayscale variance of the second re-acquisition is 88 (≥ 80), and the data is qualified. Case 4: Both parameters failed to meet the standards (visible light 75%, infrared 68%), triggering dual image re-acquisition. After three re-acquisitions, the visible light was 78% and the infrared was 69%, still failing to meet the standards, and a "data acquisition anomaly" warning was issued. This technical solution offers several key advantages: First, it accurately filters high-quality data. Through a dual-dimensional evaluation using grayscale variance and the proportion of effective temperature pixels, the proportion of high-quality data in the input feature extraction stage is increased to over 92%, effectively eliminating blurry, noisy, and invalid temperature information. Second, it reduces interference from invalid data. The filtering of high-quality data reduces the error of subsequent feature extraction by over 30%, avoiding mis-extraction of features caused by low-quality data and improving overall recognition accuracy. Third, it optimizes the re-acquisition strategy. Tiered re-acquisition (targeted re-acquisition of a single image) saves over 60% of acquisition time compared to blindly re-acquiring two images, reducing the system's data transmission and storage load. Fourth, it promptly detects acquisition anomalies. The upper limit on the number of re-acquisitions and the anomaly warning mechanism enable the failure of the acquisition unit (such as lens contamination or sensor malfunction) to be detected within 5 minutes, facilitating timely handling by maintenance personnel. Fifth, it improves system resource utilization. It avoids low-quality data occupying the computing resources of subsequent processing modules, improving the overall efficiency of feature extraction and fusion decision-making by over 25% and ensuring system real-time performance.

[0037] In some specific embodiments, when extracting visible light feature sets and infrared feature sets based on the preprocessed visual data pairs, determining whether a feature pair is redundant by calculating the correlation degree of each feature dimension in the two feature sets, calculating the discrimination score for redundant feature pairs, calculating the discrimination score of each feature after filtering, and removing weakly discriminative features according to a preset weak discrimination threshold to obtain a multimodal effective feature set, the process includes: When the correlation between two feature dimensions is greater than a preset redundancy threshold, they are determined to be a redundant feature pair. When the correlation between two feature dimensions is less than or equal to the redundancy threshold, they are determined to be non-redundant feature pairs and are directly retained. The discrimination index is obtained by calculating the mean difference and variance percentage of corresponding characteristics between healthy sheep and abnormal sheep.

[0038] It should be understood that, based on the preprocessed visual data, visible light feature sets and infrared feature sets are extracted respectively. The correlation between each feature dimension in the two feature sets is calculated to determine whether they are redundant feature pairs. Non-redundant feature pairs are directly retained. Redundant feature pairs are filtered by calculating the discriminative power (the mean difference and variance ratio of the corresponding features of healthy and abnormal sheep). Finally, a preliminary optimized feature set is obtained, which lays the foundation for the subsequent removal of weakly discriminative features and improves the effectiveness of feature representation.

[0039] The core technology of this solution is a mechanism of "dual-modal feature extraction + correlation analysis redundancy judgment + discrimination screening". In the feature extraction stage, the visible light feature set is extracted using the SIFT algorithm, including texture features (such as LBP features and gray-level co-occurrence matrix features of wool) and contour features (such as Hu moment features of the body and edge detection features), totaling 32 dimensions. The infrared feature set is extracted using histogram statistics and region segmentation, including temperature distribution features (such as average temperature, maximum temperature, temperature standard deviation, and temperature gradient), totaling 24 dimensions. The correlation degree is calculated using the Pearson correlation coefficient r, with the formula r=Cov(X,Y) / (σ). x σ y ), where Cov(X,Y) is the covariance of feature X (from the visible light feature set) and feature Y (from the infrared feature set), σ x σ y Let r be the standard deviation of the two features, and r ranges from -1 to 1. A larger absolute value indicates a stronger correlation. The preset redundancy threshold is 0.7 (determined through extensive experimental data). When |r| > 0.7, the feature pair is considered redundant; when |r| ≤ 0.7, it is considered non-redundant. The discrimination index is calculated using a comprehensive index D: "mean difference - variance ratio," calculated as D = |μ| h -μ a | / (σ h 2 +σ a 2 ), where μ h μ a σ represents the mean of the corresponding characteristics of healthy sheep and abnormal sheep, respectively. h 2 σ a 2 These represent the variances of the features corresponding to the two types of sheep. A larger D value indicates a stronger ability of the feature to distinguish between healthy and abnormal states. During the selection process, features with larger D values ​​among redundant feature pairs are retained. Feature extraction and correlation calculation are both completed in the GPU core processing module, using parallel computing to improve efficiency.

[0040] Taking the feature processing of a sheep farm as an example, the visible light feature set (32 dimensions) and infrared feature set (24 dimensions) of a sheep were extracted. The Pearson correlation coefficients for each feature dimension in the two feature sets were calculated. The correlation coefficient r = 0.82 (>0.7) between the visible light "body contour Hu moment feature X1" and the infrared "body average temperature feature Y1" was found to be redundant. The discriminant properties D1 = 0.65 for X1 and D2 = 0.78 for Y1 were calculated, so Y1 was retained and X1 was removed. Simultaneously, the correlation coefficient r = 0.58 (≤0.7) between the visible light "wool LBP feature X2" and the infrared "temperature gradient feature Y2" was found to be non-redundant and was directly retained. After correlation analysis and discriminant property selection, 22 dimensions of the 32-dimensional visible light feature set and 18 dimensions of the 24-dimensional infrared feature set were retained, resulting in a preliminary optimized feature set of 40 dimensions. The technical solution offers significant advantages: First, it accurately eliminates redundant features. Through Pearson correlation coefficient analysis, the proportion of redundant features is reduced from the initial 37.5% to 12%, reducing feature dimensions while avoiding information duplication. Second, it enhances feature discrimination capabilities. High-discrimination features are retained through discrimination screening, improving the feature set's ability to distinguish between healthy and abnormal sheep by more than 35%, laying the foundation for subsequent improvements in recognition accuracy. Third, it improves feature processing efficiency. The feature dimensions are reduced by 40%, shortening the time required for subsequent feature calculations (such as normalization and fusion) by more than 50%, ensuring system real-time performance. Fourth, it reduces interference caused by data redundancy. The elimination of redundant features reduces the impact of irrelevant information on fusion decisions, improving the training efficiency of the fusion model by more than 25% and enhancing generalization ability. Fifth, it adapts to multimodal fusion requirements. The retained non-redundant features and high-discrimination features cover the core information of visible light and infrared, achieving complementarity between the two types of information and improving the synergistic effect of multimodal fusion.

[0041] In some specific embodiments, when extracting visible light feature sets and infrared feature sets based on the preprocessed visual data pairs, determining whether a feature pair is redundant by calculating the correlation degree of each feature dimension in the two feature sets, calculating the discrimination score for redundant feature pairs, calculating the discrimination score of each feature after filtering, and removing weakly discriminative features according to a preset weak discrimination threshold to obtain a multimodal effective feature set, the process includes: A weak discrimination threshold is preset, the discrimination of each feature after filtering is calculated, and it is determined whether the discrimination of each feature is greater than or equal to the weak discrimination threshold. When the feature discrimination is less than the weak discrimination threshold, it is determined to be a weak discrimination feature and is removed. When the feature discrimination is greater than or equal to the weak discrimination threshold, it is determined to be a valid feature and retained in the multimodal valid feature set; After the elimination is completed, it is determined whether the dimension of the multimodal effective feature set is greater than or equal to the preset minimum dimension requirement. If it is not met, the features are extracted again.

[0042] It should be understood that a weak discrimination threshold is preset, the discrimination of each feature after redundancy screening is calculated, weak discrimination features with discrimination less than the threshold are removed, and effective features are retained to form a multimodal effective feature set. After the removal is completed, it is judged whether the dimension of the effective feature set meets the preset minimum dimension requirement. If it does not meet the requirement, features are re-extracted to ensure that the final effective feature set has sufficient dimension and strong discrimination ability to ensure the accuracy of subsequent fusion recognition.

[0043] The core technology of this solution is a closed-loop mechanism of "weak discrimination feature determination and elimination + dimensional threshold verification + feature re-extraction triggering". The weak discrimination threshold is determined based on a large number of feature samples of healthy and abnormal sheep. By calculating the discrimination of all sample features, the 25th percentile of the discrimination distribution is taken as the weak discrimination threshold (set to 0.3 in this embodiment). When the discrimination of a feature D < 0.3, it is determined to be a weak discrimination feature. Such features have extremely weak ability to distinguish between healthy and abnormal states, which will increase the computational load of the system and interfere with the recognition results, and must be eliminated. The preset minimum dimension requirement of the multimodal effective feature set is determined based on the input requirements of the fusion model. Through experimental verification, when the feature dimension is ≥ 10, the fusion model can maintain high recognition accuracy. Therefore, the minimum dimension threshold is set to 10. In the dimensionality verification stage, if the effective feature set after removing weakly discriminative features has a dimension ≥ 10, then this feature set is output for subsequent fusion. If the dimension is < 10, it indicates that there are omissions or deviations in the current feature extraction, triggering a re-extraction process. During re-extraction, the feature extraction parameters (such as the feature point detection threshold of the SIFT algorithm and the region segmentation accuracy of infrared features) are adjusted to ensure that a sufficient number of effective features are extracted. The maximum number of re-extractions is 3. If the dimensionality requirement is still not met after 3 extractions, a "feature extraction anomaly" warning is issued, reminding operations and maintenance personnel to check the preprocessing quality and feature extraction algorithm parameters.

[0044] Taking the process of weakly discriminative feature removal and dimensionality verification in a sheep farm as an example, after redundancy screening, a 40-dimensional feature set is obtained. The discriminative power of each feature is calculated, and it is found that the discriminative powers of 6 features are 0.22, 0.25, 0.18, 0.28, 0.21, and 0.26 (all < the weak discriminative threshold of 0.3), which are identified as weakly discriminative features and removed. Finally, a 34-dimensional multimodal effective feature set is obtained, with a dimension of 34 ≥ 10, which meets the minimum dimensionality requirement and is output for subsequent fusion. If another set of data yields a 12-dimensional feature set after redundancy screening, removing 2 weakly discriminative features results in a 10-dimensional effective feature set, which just meets the minimum dimensionality requirement. If the removal results in a 9-dimensional feature set (< 10), the re-extraction process is triggered, and the feature point detection threshold of the SIFT algorithm is adjusted from 0.03 to 0.02. After re-extraction, a 15-dimensional feature set after redundancy screening is obtained, and removing 3 weakly discriminative features results in a 12-dimensional effective feature set, which meets the dimensionality requirement. The technical solution offers several key advantages: First, it further optimizes the quality of the feature set. The removal of weakly discriminative features increases the average discriminative power of the effective feature set by over 45%, preventing interference from weakly discriminative features in the recognition results. Second, it ensures the effectiveness of the feature set dimensions. The minimum dimension verification mechanism ensures that the effective feature set has sufficient information carrying capacity, maintaining the recognition accuracy of the fusion model above 90% and preventing performance degradation due to insufficient feature dimensions. Third, it improves the efficiency of the fusion model. The effective feature set dimensions are streamlined (reduced from 40 to 34 dimensions in this embodiment), shortening the inference time of the fusion model by over 30% and ensuring real-time system response. Fourth, it promptly detects feature extraction problems. The re-extraction mechanism and anomaly warning allow deviations in the feature extraction process to be corrected quickly, reducing system failures caused by feature extraction failures. Fifth, it enhances system robustness. By adjusting parameters through feature re-extraction, the feature extraction process can adapt to sheep in different states (such as different breeds and different growth stages), improving the system's adaptability to diverse farming scenarios.

[0045] In some specific embodiments, the process of normalizing visible light effective features and infrared effective features and setting initial fusion weights to obtain initial fusion features, introducing a feature consistency coefficient to judge the consistency of prediction results between the two feature modes, dynamically adjusting the fusion weights and updating the fusion features according to the range of the feature consistency coefficient, combining hierarchical decision fusion to obtain the final prediction result, and then determining the sheep's health and emotional state according to a preset confidence threshold, and classifying the warning level and outputting corresponding warning information based on the duration of the state, includes: A high consistency threshold and a medium consistency threshold are preset. It is determined whether the feature consistency coefficient is greater than or equal to the high consistency threshold. When the feature consistency coefficient is greater than or equal to the high consistency threshold, the initial fusion weight is maintained. When the feature consistency coefficient is less than the high consistency threshold and greater than or equal to the medium consistency threshold, the historical prediction accuracy of the two feature modalities is calculated, and the fusion weight is adjusted according to the accuracy ratio. When the feature consistency coefficient is less than the consistency threshold, the discriminative power of the two feature modes is determined, and the weight corresponding to the feature mode with higher discriminative power is greater than the weight corresponding to the feature mode with lower discriminative power. When two feature modes have the same discriminative power, the weights corresponding to the two feature modes are the same.

[0046] It should be understood that after normalizing the visible light effective features and infrared effective features, initial fusion weights are set to obtain initial fusion features. A feature consistency coefficient is introduced to judge the consistency of the prediction results of the two feature modes. High consistency threshold and medium consistency threshold are preset. According to the range of the consistency coefficient, strategies of "maintaining the initial weight", "adjusting the weight according to the historical accuracy" and "adjusting the weight according to the discrimination" are adopted respectively. When the discrimination of the two feature modes is the same, equal weights are adopted to ensure that the fusion weights are adapted to the difference in prediction performance of the two feature modes and improve the reliability of the fusion features.

[0047] The core technology of this solution is a mechanism of "feature normalization + initial weight setting + consistency coefficient determination + dynamic weight adjustment". Feature normalization uses the min-max normalization method, mapping the effective feature value range to [0,1]. The calculation formula is x'=(xx... min ) / (x max -x min ), where x is the original eigenvalue, x min x max The minimum and maximum values ​​of the feature are set to ensure that the two modal features are on the same order of magnitude, avoiding weight bias caused by differences in magnitude. The initial fusion weights are determined based on expert experience and experimental data, taking into account the average discriminative power and historical prediction accuracy of the two modal features. In this embodiment, the initial weights are set to 0.4 for visible light and 0.6 for infrared. The feature consistency coefficient is calculated using cosine similarity s, s=Cos(θ)=A·B / (|A||B|), where A and B are the prediction result vectors obtained by inputting the effective features of visible light and infrared into the same basic prediction model (logistic regression model), respectively. The value range of s is [0,1], and the larger the s is, the higher the consistency of the prediction results of the two modalities. The preset high consistency threshold s1=0.9 and medium consistency threshold s2=0.7 form three consistency intervals: s≥s1 (high consistency), s2≤s<s1 (medium consistency), and s<s2 (low consistency). Historical prediction accuracy is calculated by statistically analyzing the two-modal prediction accuracy of the past 1000 data sets (Accᵥ for visible light prediction accuracy and Accᵢ for infrared prediction accuracy), with weights adjusted according to ω. v =Accv / (Acc v +Acc v ), ω i =Acc v / (Acc v +Acc i ) Calculate; for low consistency, compare the average discriminant (D) of the effective feature sets of the two modalities. v D represents the average resolution of visible light. i (For infrared average discrimination), mode weights are higher when D is larger. v =D i ω v =ω i =0.5.

[0048] Taking the feature fusion weight adjustment process of a sheep farm as an example, Case 1: After min-max normalization, the visible light effective features (17 dimensions) and infrared effective features (17 dimensions) have an initial fusion weight of 0.4-0.6. These are input into the logistic regression model to obtain prediction result vectors A and B. The cosine similarity s is calculated to be 0.92 ≥ 0.9 (high consistency). The initial weights are maintained, resulting in the initial fused features. Case 2: s is calculated to be 0.82 (between 0.7 and 0.9), and the historical prediction accuracy Acc is statistically analyzed. v =85%, Acc i =90%, calculate the adjusted weight ω v =85 / (85+90)=0.485、ω i =90 / (85+90)=0.515, update the fusion weight to 0.485-0.515. Case 3: Calculate s=0.65<0.7 (low consistency), calculate the average discrimination D of the two modalities. v =0.58, D i =0.65, D i >D v The weights are adjusted to 0.45-0.55; if the other set of data D v =0.62、D i=0.62, with the weight adjusted to 0.5-0.5. The technical effects of this solution are significant: First, it improves the reliability of fused features. Dynamic weight adjustment allows the fused features to fully utilize the advantages of both modal features, improving feature representation capabilities by more than 30% compared to fixed-weight fusion. Second, it adapts to the differences in two-modal predictions. Differentiated weight strategies for different consistency intervals ensure the fusion model remains stable when the two modal predictions are consistent, and prioritizes the information from the superior modality when they are inconsistent, improving recognition accuracy by 15-20%. Third, it enhances the adaptability of the fusion system. The introduction of historical accuracy and discriminative power allows weight adjustment to adapt to changes in feature performance under different farming scenarios and sheep conditions, improving system robustness by more than 25%. Fourth, it reduces prediction bias. Dynamic weight optimization reduces the impact of single-modal performance fluctuations on the fusion results, reducing the standard deviation of fusion predictions by more than 40%. Fifth, it improves fusion efficiency. Based on a simple and effective consistency coefficient and weight calculation method, the weight adjustment time is controlled within 0.1 seconds, ensuring the system's real-time fusion requirements.

[0049] In some specific embodiments, the process of normalizing visible light effective features and infrared effective features and setting initial fusion weights to obtain initial fusion features, judging the consistency of prediction results between two feature modes by feature consistency coefficient, dynamically adjusting fusion weights and updating fusion features according to the range of feature consistency coefficient, obtaining the final prediction result by combining hierarchical decision fusion, determining the sheep's health and emotional state according to a preset confidence threshold, classifying warning levels based on the duration of the state and outputting corresponding warning information includes: The updated fused features are input into the machine learning model to obtain preliminary prediction results. At the same time, the visible light effective features and infrared effective features are input into the same machine learning model to obtain two single-modal prediction results. The weight of the preliminary prediction result is greater than the weight of the two single-modality prediction results; When the deviation of the prediction results for a preset number of consecutive predictions is less than a preset deviation threshold, it is determined to be the final prediction result; If the deviation of the prediction results for a preset number of consecutive predictions is greater than or equal to the deviation threshold, then the fusion calculation is performed again.

[0050] It should be understood that the updated fused features are input into the machine learning model to obtain preliminary prediction results, while the effective features of the two single modalities are input into the same model to obtain two single modal prediction results. The weight of the preliminary prediction results is set higher than the weight of the single modal prediction results. By judging whether the deviation of the prediction results of a consecutive preset number of times is less than the preset deviation threshold, the final prediction result is determined or a re-fusion calculation is triggered to ensure the stability and reliability of the final prediction result.

[0051] The machine learning model used is a random forest model (100 decision trees, maximum depth 10), which has strong feature fitting and generalization capabilities and is suitable for prediction tasks with multi-dimensional features. The weight allocation for hierarchical decision fusion is set based on the information integrity of the prediction results. The initial prediction result (obtained by inputting fused features into the model) has a weight ω1=0.6, and the two single-modal prediction results (obtained by inputting visible light features) have a weight P... v Infrared feature input yields P i The weights ω2=ω3=0.2, and the total weight sum is 1. The preset number of consecutive predictions is set to 3 (experimental verification shows that 3 consecutive predictions can effectively eliminate random bias). The prediction result deviation is calculated using the mean squared error (MSE), MSE=Σ(y k -ŷ k ) 2 / 3, where y k The prediction result after the k-th hierarchical fusion is ŷ k The mean of the three prediction results is used, with a preset deviation threshold of 0.05 (corresponding to a prediction probability deviation of 5%). If MSE < 0.05, it indicates that the continuous prediction results are stable, and the mean of the three prediction results is determined as the final prediction result. If MSE ≥ 0.05, it indicates that the prediction results fluctuate greatly, triggering the re-fusion calculation process. During re-fusion, the feature consistency coefficient is recalculated, the fusion weights are adjusted, and hierarchical decision fusion is performed again. The maximum number of re-fusions is 2. If the condition is still not met after 2 re-fusions, a "prediction anomaly" warning is issued.

[0052] Taking the hierarchical decision fusion process of a sheep farm as an example, the updated fusion features are input into the random forest model to obtain the preliminary prediction result P1=0.92 (probability of healthy state), and the visible light single-modal prediction result P v =0.88, Infrared single-mode prediction result P i =0.94, the result of the first layered fusion is y1=0.6×0.92+0.2×0.88+0.2×0.94=0.92; the result of the second fusion is y2=0.93; the result of the third fusion is y3=0.91, and the mean of the three results is ŷ=0.92; Mean squared error MSE = [(0.92-0.92)] 2 +(0.93-0.92) 2 +(0.91-0.92) 2 Since ] / 3≈0.000067<0.05, the final prediction result is determined to be 0.92, corresponding to the health status label. If the results of the three stratified fusions of another set of data are 0.85, 0.95, and 0.78 respectively, with a mean ŷ=0.86, then MSE=[(0.85-0.86)] 2 +(0.95-0.86) 2+(0.78-0.86) 2 If ] / 3≈0.0053≥0.05, a re-fusion calculation is triggered. The feature consistency coefficient s=0.68 is recalculated, the fusion weight is adjusted to 0.45-0.55, and the layered fusion is performed again to obtain 3 results: 0.88, 0.90, and 0.89. MSE≈0.0001<0.05, and the final prediction result is determined to be 0.89. The technical solution offers several key advantages: First, it enhances the reliability of prediction results. Layered decision fusion, combined with multimodal prediction information, improves accuracy by over 10% compared to single-feature fusion prediction. Continuous deviation verification keeps the fluctuation error of the final prediction result within 5%. Second, it reduces the impact of random errors. Three consecutive predictions and mean calculation effectively eliminate random deviations from single predictions, improving the stability of prediction results by over 40%. Third, it ensures prediction accuracy. The re-fusion mechanism allows for timely adjustment of the fusion strategy when prediction results fluctuate significantly, avoiding misjudgments caused by a single prediction deviation, reducing the misjudgment rate to below 3%. Fourth, it enhances the robustness of the prediction system. The combination of layered weight allocation and deviation verification enables the system to adapt to prediction needs under different feature qualities, maintaining high prediction accuracy even with slight noise in the features. Fifth, it promptly detects prediction anomalies. The re-fusion upper limit and early warning mechanism ensure timely detection of prediction model faults or parameter deviations, guaranteeing the stable operation of the system's prediction function.

[0053] In some specific embodiments, the process of normalizing visible light effective features and infrared effective features and setting initial fusion weights to obtain initial fusion features, judging the consistency of prediction results between two feature modes by feature consistency coefficient, dynamically adjusting fusion weights and updating fusion features according to the range of feature consistency coefficient, obtaining the final prediction result by combining hierarchical decision fusion, determining the sheep's health and emotional state according to a preset confidence threshold, classifying warning levels based on the duration of the state and outputting corresponding warning information includes: A high confidence threshold and a suspected confidence threshold are preset. It is determined whether the confidence of the final prediction result is greater than or equal to the high confidence threshold. When the confidence of the final prediction result is greater than or equal to the high confidence threshold, the corresponding sheep health and emotional state labels are output. If the confidence level of the final prediction result is greater than or equal to the suspected confidence level threshold and less than the high confidence level threshold, then a second data collection is performed. When the confidence level of the final prediction result is greater than or equal to the suspected confidence threshold and less than the high confidence threshold after the second data collection, a second warning signal is issued. If the confidence level of the final prediction result is less than the suspected confidence level threshold, the data is determined to be abnormal and the data is collected again. Preset short-term duration thresholds and long-term duration thresholds. Based on the type of abnormal state and whether the duration is greater than or equal to the corresponding threshold, the system determines the level of warning as Level 1, Level 2, or Level 3, respectively.

[0054] It should be understood that by setting high confidence thresholds and suspected confidence thresholds, the status label is determined based on the confidence of the final prediction result, triggering secondary data collection or data anomaly processing; by setting short-term and long-term duration thresholds, the first, second, and third level early warnings are divided according to the type and duration of the abnormal status, and corresponding early warning information is output to ensure accurate status determination, timely early warning and reasonable classification, and to adapt to the differentiated handling needs of aquaculture personnel.

[0055] The confidence score is the probability value of the final prediction result, corresponding to different health and emotional state labels (e.g., healthy state: confidence score ≥ 0.9; mild stress: 0.7 ≤ confidence score < 0.9; severe stress: confidence score < 0.7; illness: confidence score < 0.7 and abnormal infrared temperature characteristics). A preset high confidence threshold of 0.9 and a suspected confidence threshold of 0.7 are used. A confidence score ≥ 0.9 directly outputs the corresponding state label; a confidence score ≤ 0.7 < 0.9 triggers a second data collection (data is recollected and prediction is completed after a 5-minute interval). If the data still falls within this range after the second collection, a second warning signal ("suspected anomaly") is issued; a confidence score < 0.7 is considered abnormal, triggering a recollection of data. The duration of the state is tracked in real time by a timer, starting from the first determination of an abnormal state (mild stress, severe stress, illness). Preset short-term duration thresholds are 2 hours and long-term duration thresholds are 8 hours. Warning levels are determined based on the type of abnormal state: Mild stress lasting ≥2 hours and <8 hours is a Level 1 warning (requiring observation); Mild stress lasting ≥8 hours or severe stress lasting ≥2 hours and <8 hours is a Level 2 warning (requiring intervention); Severe stress lasting ≥8 hours or illness lasting ≥2 hours is a Level 3 warning (requiring emergency response). Warning information includes the abnormal state type, duration, warning level, corresponding sheep number and location, and is output through both a local audible and visual alarm (different levels correspond to different alarm sounds and lights: Level 1 flashing green, Level 2 solid yellow, Level 3 flashing red + high-decibel alarm) and a remote monitoring platform (SMS push + pop-up notification).

[0056] Taking the status assessment and graded early warning process of a sheep farm as an example: Case 1: The final prediction result confidence level = 0.93 ≥ 0.9, and the "healthy status" label is output. Case 2: Confidence level = 0.82 (0.7 ≤ 0.82 < 0.9), triggering a second data collection. After 5 minutes, the prediction result is 0.85, still in the suspected range, and a second early warning signal ("suspected mild stress") is issued. Case 3: The initial judgment is "mild stress" (confidence level = 0.8), the timer starts, and after 2.5 hours (≥ 2 hours and < 8 hours), a first-level early warning is triggered. The local green light flashes, and the remote platform pushes the message "Sheep No. 32 is suspected of mild stress, lasting for 2.5 hours, observation is recommended." Case 4: The "severe stress" status lasts for 3 hours (≥ 2 hours and < 8 hours), triggering a second-level early warning. The local yellow light stays on, and the message "Sheep No. 15 is severely stressed, lasting for 3 hours, immediate intervention is recommended." Case 5: The "sick" status lasted for 2 hours, triggering a level 3 alert. The local red light flashed and a 90dB alarm sounded. The message "Sheep No. 27 is suspected of being sick, lasting for 2 hours. Emergency action required!" was pushed to the police. The technical solutions offer significant advantages: First, they accurately determine the sheep's condition. Confidence grading and secondary data collection and verification improve the accuracy of condition determination to over 96%, reducing missed and false alarms. Second, they provide timely anomaly warnings. Secondary warning signals allow suspected anomalies to be addressed early, preventing the condition from worsening and reducing the warning response time to less than 10 minutes. Third, tiered warnings adapt to different handling needs. Different warning levels correspond to different handling strategies, allowing farmers to allocate their time effectively. Level 1 warnings require only observation, reducing maintenance costs, while Level 3 warnings require emergency handling, minimizing sheep casualties. Fourth, they enhance the precision of livestock management. Warning information includes sheep numbers and locations, enabling farmers to quickly locate target sheep and improving handling efficiency by over 60%. Fifth, they reduce livestock risks. Tiered warnings and timely handling reduce sheep mortality due to abnormal conditions by over 5%, and the recovery rate after timely intervention for mild stress increases by over 80%, significantly improving livestock profitability.

[0057] See Figure 3 As shown in the second embodiment, a multi-source information fusion intelligent identification system for sheep health monitoring according to an embodiment of this application includes: The acquisition module is configured to deploy a visible light acquisition unit and an infrared acquisition unit for acquiring visual data of sheep in the target sheep breeding area. The visible light acquisition unit and the infrared acquisition unit are synchronized. Based on the verification result, it is determined whether to perform unit synchronization calibration. The visible light image and the infrared image that pass the verification are paired to form a visual data pair. The verification module is configured to preprocess the visual data pairs, preset a visible light image sharpness threshold and an infrared image temperature validity threshold, perform quality verification on the preprocessed visual data pairs according to the preset thresholds, and determine whether to re-acquire data based on the verification results. The filtering module is configured to extract visible light feature sets and infrared feature sets based on the preprocessed visual data pairs, determine whether the feature pairs are redundant by calculating the correlation degree of each feature dimension in the two feature sets, calculate the discrimination degree of redundant feature pairs and filter them, calculate the discrimination degree of each feature after filtering, and remove weakly discriminative features according to a preset weak discrimination threshold to obtain a multimodal effective feature set. The early warning module is configured to normalize the effective features of visible light and infrared light and set initial fusion weights to obtain initial fusion features. It judges the consistency of the prediction results of the two feature modes by the feature consistency coefficient, dynamically adjusts the fusion weights and updates the fusion features according to the range of the feature consistency coefficient, and obtains the final prediction result by combining hierarchical decision fusion. Then, it determines the health and emotional state of the sheep according to the preset confidence threshold, divides the early warning level according to the state duration and outputs the corresponding early warning information.

[0058] It should be understood that this system adopts a modular architecture design. Each module interacts with and transmits commands to the core control unit (CPU+GPU architecture, CPU is Intel Core i7-12700H, GPU is NVIDIA Jetson AGXXavier) via industrial Ethernet. The core control unit is responsible for module scheduling and overall process control. The acquisition module consists of a visible light acquisition unit (4 sets of high-definition industrial cameras, deployed at the four corners of the breeding area, covering the entire area), an infrared acquisition unit (4 sets of uncooled infrared thermal imagers, deployed one-to-one with the visible light cameras), a synchronization trigger module, and a speed monitoring sensor. Its core function is to complete data acquisition and synchronization pairing. The synchronization trigger module receives commands from the core control unit and uniformly controls the acquisition timing of the two acquisition units. The speed monitoring sensor provides data support for time difference threshold calculation. The verification module includes a preprocessing submodule and a quality judgment submodule. The preprocessing submodule integrates Gaussian filtering, adaptive histogram equalization, median filtering, and pseudo-color mapping algorithms, and dynamically adjusts parameters upon receiving data from environmental sensors (light, temperature). The quality judgment submodule calculates grayscale variance, effective temperature pixel ratio, and threshold comparison, triggering a re-acquisition command. The filtering module includes a feature extraction submodule, a redundancy removal submodule, and a weak discrimination removal submodule. The feature extraction submodule integrates the SIFT algorithm and the infrared feature statistics algorithm; the redundancy removal submodule calculates the Pearson correlation coefficient and determines redundancy; the weak discrimination removal submodule calculates the discrimination score and verifies the dimension, triggering feature re-extraction. The early warning module includes a fusion submodule, a decision submodule, and an early warning submodule. The fusion submodule performs feature normalization and dynamic weight adjustment; the decision submodule integrates the random forest model and the hierarchical decision fusion algorithm to determine the final prediction result; the early warning submodule performs duration statistics, early warning level classification, and early warning information output. The system also includes a power supply module (220V AC input, DC 12V output, power supply stability ±0.1V), a storage module (1TB SSD, storing raw data, prediction results, and early warning information), and a communication module (5G + Ethernet, ensuring remote data transmission).

[0059] Taking the system operation of a large-scale sheep farm (500 sheep, 1000㎡ breeding area) as an example, after the core control unit is started, the scheduling and acquisition module starts working. Four sets of visible light-infrared acquisition units collect data once every 10 seconds. The synchronous trigger module controls the acquisition sequence. The speed monitoring sensor collects the sheep's walking speed and calculates the time difference threshold. After synchronous verification, a valid data pair is formed. The data is transmitted to the verification module. The preprocessing submodule adjusts the parameters according to the data from the light sensor (8000 lux on a sunny day) and the temperature sensor (32℃) to complete the preprocessing. The quality judgment submodule calculates that the visible light clarity = 95 and the infrared temperature validity = 78%, and judges it as high-quality data. The screening module extracts 32-dimensional visible light features and 24-dimensional infrared features. The redundancy removal submodule removes 12 pairs of redundant features, and the weak differentiation removal submodule removes 6 features.

[0060] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0061] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0062] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0063] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-source information fusion intelligent identification method for sheep health monitoring, characterized in that, Includes the following steps: A visible light acquisition unit and an infrared acquisition unit are deployed to collect visual data of sheep in the target sheep breeding area. The visible light acquisition unit and the infrared acquisition unit are synchronized. Based on the verification result, it is determined whether to perform unit synchronization calibration. The visible light image and the infrared image that pass the verification are paired to form a visual data pair. The visual data pairs are preprocessed, and a visible light image sharpness threshold and an infrared image temperature validity threshold are preset. The quality of the preprocessed visual data pairs is checked according to the preset thresholds, and it is determined whether to re-acquire data based on the check results. Based on the preprocessed visual data pairs, visible light feature sets and infrared feature sets are extracted respectively. The correlation degree of each feature dimension in the two feature sets is calculated to determine whether the feature pairs are redundant. The redundancy feature pairs are filtered by calculating the discrimination degree, and the discrimination degree of each feature after filtering is calculated. Weak discrimination features are removed according to the preset weak discrimination threshold to obtain the multimodal effective feature set. The effective features of visible light and infrared light are normalized and initial fusion weights are set to obtain initial fusion features. The consistency of the prediction results of the two feature modes is judged by the feature consistency coefficient. The fusion weights are dynamically adjusted and the fusion features are updated according to the range of the feature consistency coefficient. The final prediction result is obtained by combining hierarchical decision fusion. The health and emotional state of sheep are determined according to the preset confidence threshold. The warning level is divided according to the duration of the state and the corresponding warning information is output.

2. The multi-source information fusion intelligent identification method for sheep health monitoring according to claim 1, characterized in that, The deployment of visible light acquisition units and infrared acquisition units for collecting visual data of sheep in the target sheep breeding area, performing synchronization verification on the visible light acquisition units and infrared acquisition units, determining whether to perform unit synchronization calibration based on the verification results, and pairing the verified visible light images with infrared images to form visual data pairs includes: Based on the sheep's walking speed, a time difference threshold for acquisition is determined, and it is determined whether the acquisition time difference between the visible light image and the infrared image is less than or equal to the acquisition time difference threshold. When the acquisition time difference between the visible light image and the infrared image is greater than the acquisition time difference threshold, a synchronization check is performed, and a synchronization check is performed again after calibration. If the acquisition time difference between the visible light image and the infrared image after secondary calibration is greater than the acquisition time difference threshold, a first warning signal will be issued. If the acquisition time difference between the visible light image and the infrared image after secondary calibration is less than or equal to the acquisition time difference threshold, then the data pair is determined to be valid.

3. The multi-source information fusion intelligent identification method for sheep health monitoring according to claim 2, characterized in that, The preprocessing of the visual data pairs includes: Gaussian filtering is applied to the visible light image to remove environmental noise, and adaptive histogram equalization is then used to enhance the contrast between fur texture and body contour. The infrared image is filtered by median filtering to remove thermal noise, and then the body temperature distribution is transformed into a visual feature through pseudo-color mapping. At the same time, the body temperature data matrix corresponding to the infrared image is extracted. The filtering parameters of the visible light image preprocessing are dynamically adjusted according to the ambient brightness of the target sheep breeding area, and the noise removal intensity of the infrared image preprocessing is dynamically adjusted according to the ambient temperature of the target sheep breeding area.

4. The multi-source information fusion intelligent identification method for sheep health monitoring according to claim 3, characterized in that, The process of preprocessing the visual data pairs, setting a visible light image sharpness threshold and an infrared image temperature validity threshold, performing quality checks on the preprocessed visual data pairs based on the preset thresholds, and determining whether to re-acquire data based on the check results includes: The sharpness of the visible light image is calculated using the gray-scale variance method, and a threshold for the sharpness of the visible light image is preset. The temperature validity of the infrared image is calculated based on the percentage of effective temperature pixels, and a preset threshold for the temperature validity of the infrared image is established. When the visible light image sharpness in the preprocessed visual data pair is greater than or equal to the visible light image sharpness threshold, and the infrared image temperature validity is greater than or equal to the infrared image temperature validity threshold, it is determined to be a high-quality image; When the visible light image sharpness in the preprocessed visual data pair is greater than or equal to the visible light image sharpness threshold, and the infrared image temperature validity is less than the infrared image temperature validity threshold, the infrared image is reacquired. When the visible light image sharpness in the preprocessed visual data pair is less than the visible light image sharpness threshold, and the infrared image temperature validity is greater than or equal to the infrared image temperature validity threshold, the visible light image is re-acquired; When the visible light image sharpness in the preprocessed visual data pair is less than the visible light image sharpness threshold, and the infrared image temperature validity is less than the infrared image temperature validity threshold, the visible light image and the infrared image are reacquired.

5. The multi-source information fusion intelligent identification method for sheep health monitoring according to claim 4, characterized in that, The process of extracting visible light and infrared feature sets from the preprocessed visual data pairs, determining whether a feature pair is redundant by calculating the correlation of each feature dimension in the two feature sets, calculating the discrimination of redundant feature pairs, calculating the discrimination of each feature after filtering, and removing weakly discriminative features according to a preset weak discrimination threshold to obtain a multimodal effective feature set includes: When the correlation between two feature dimensions is greater than a preset redundancy threshold, they are determined to be a redundant feature pair. When the correlation between two feature dimensions is less than or equal to the redundancy threshold, they are determined to be non-redundant feature pairs and are directly retained. The discrimination index is obtained by calculating the mean difference and variance percentage of corresponding characteristics between healthy sheep and abnormal sheep.

6. The multi-source information fusion intelligent identification method for sheep health monitoring according to claim 5, characterized in that, The process of extracting visible light and infrared feature sets from the preprocessed visual data pairs, determining whether a feature pair is redundant by calculating the correlation of each feature dimension in the two feature sets, calculating the discrimination of redundant feature pairs, calculating the discrimination of each feature after filtering, and removing weakly discriminative features according to a preset weak discrimination threshold to obtain a multimodal effective feature set includes: A weak discrimination threshold is preset, the discrimination of each feature after filtering is calculated, and it is determined whether the discrimination of each feature is greater than or equal to the weak discrimination threshold. When the feature discrimination is less than the weak discrimination threshold, it is determined to be a weak discrimination feature and is removed. When the feature discrimination is greater than or equal to the weak discrimination threshold, it is determined to be a valid feature and retained in the multimodal valid feature set; After the elimination is completed, it is determined whether the dimension of the multimodal effective feature set is greater than or equal to the preset minimum dimension requirement. If it is not met, the features are extracted again.

7. The multi-source information fusion intelligent identification method for sheep health monitoring according to claim 6, characterized in that, The process of normalizing visible light and infrared effective features and setting initial fusion weights to obtain initial fusion features, introducing a feature consistency coefficient to judge the consistency of prediction results between the two feature modes, dynamically adjusting the fusion weights and updating the fusion features according to the range of the feature consistency coefficient, combining hierarchical decision fusion to obtain the final prediction result, and then determining the sheep's health and emotional state according to a preset confidence threshold, and classifying the warning level and outputting corresponding warning information based on the state duration includes: A high consistency threshold and a medium consistency threshold are preset. It is determined whether the feature consistency coefficient is greater than or equal to the high consistency threshold. When the feature consistency coefficient is greater than or equal to the high consistency threshold, the initial fusion weight is maintained. When the feature consistency coefficient is less than the high consistency threshold and greater than or equal to the medium consistency threshold, the historical prediction accuracy of the two feature modalities is calculated, and the fusion weight is adjusted according to the accuracy ratio. When the feature consistency coefficient is less than the consistency threshold, the discriminative power of the two feature modes is determined, and the weight corresponding to the feature mode with higher discriminative power is greater than the weight corresponding to the feature mode with lower discriminative power. When two feature modes have the same discriminative power, the weights corresponding to the two feature modes are the same.

8. The multi-source information fusion intelligent identification method for sheep health monitoring according to claim 7, characterized in that, The process of normalizing visible light and infrared effective features and setting initial fusion weights to obtain initial fusion features, judging the consistency of prediction results between two feature modes through feature consistency coefficients, dynamically adjusting fusion weights and updating fusion features based on the range of the feature consistency coefficients, obtaining the final prediction result through hierarchical decision fusion, and then determining the sheep's health and emotional state based on a preset confidence threshold, classifying warning levels based on the duration of the state, and outputting corresponding warning information includes: The updated fused features are input into the machine learning model to obtain preliminary prediction results. At the same time, the visible light effective features and infrared effective features are input into the same machine learning model to obtain two single-modal prediction results. The weight of the preliminary prediction result is greater than the weight of the two single-modality prediction results; When the deviation of the prediction results for a preset number of consecutive predictions is less than a preset deviation threshold, it is determined to be the final prediction result; If the deviation of the prediction results for a preset number of consecutive predictions is greater than or equal to the deviation threshold, then the fusion calculation is performed again.

9. The multi-source information fusion intelligent identification method for sheep health monitoring according to claim 8, characterized in that, The process of normalizing visible light and infrared effective features and setting initial fusion weights to obtain initial fusion features, judging the consistency of prediction results between two feature modes through feature consistency coefficients, dynamically adjusting fusion weights and updating fusion features based on the range of the feature consistency coefficients, obtaining the final prediction result through hierarchical decision fusion, and then determining the sheep's health and emotional state based on a preset confidence threshold, classifying warning levels based on the duration of the state, and outputting corresponding warning information includes: A high confidence threshold and a suspected confidence threshold are preset. It is determined whether the confidence of the final prediction result is greater than or equal to the high confidence threshold. When the confidence of the final prediction result is greater than or equal to the high confidence threshold, the corresponding sheep health and emotional state labels are output. If the confidence level of the final prediction result is greater than or equal to the suspected confidence level threshold and less than the high confidence level threshold, then a second data collection is performed. When the confidence level of the final prediction result is greater than or equal to the suspected confidence threshold and less than the high confidence threshold after the second data collection, a second warning signal is issued. If the confidence level of the final prediction result is less than the suspected confidence level threshold, the data is determined to be abnormal and the data is collected again. Preset short-term duration thresholds and long-term duration thresholds. Based on the type of abnormal state and whether the duration is greater than or equal to the corresponding threshold, the system determines the level of warning as Level 1, Level 2, or Level 3, respectively.

10. A multi-source information fusion intelligent identification system for sheep health monitoring, characterized in that, The multi-source information fusion intelligent identification method for sheep health monitoring, as described in any one of claims 1 to 9, includes: The acquisition module is configured to deploy a visible light acquisition unit and an infrared acquisition unit for acquiring visual data of sheep in the target sheep breeding area. The visible light acquisition unit and the infrared acquisition unit are synchronized. Based on the verification result, it is determined whether to perform unit synchronization calibration. The visible light image and the infrared image that pass the verification are paired to form a visual data pair. The verification module is configured to preprocess the visual data pairs, preset a visible light image sharpness threshold and an infrared image temperature validity threshold, perform quality verification on the preprocessed visual data pairs according to the preset thresholds, and determine whether to re-acquire data based on the verification results. The filtering module is configured to extract visible light feature sets and infrared feature sets based on the preprocessed visual data pairs, determine whether the feature pairs are redundant by calculating the correlation degree of each feature dimension in the two feature sets, calculate the discrimination degree of redundant feature pairs and filter them, calculate the discrimination degree of each feature after filtering, and remove weakly discriminative features according to a preset weak discrimination threshold to obtain a multimodal effective feature set. The early warning module is configured to normalize the effective features of visible light and infrared light and set initial fusion weights to obtain initial fusion features. It judges the consistency of the prediction results of the two feature modes by the feature consistency coefficient, dynamically adjusts the fusion weights and updates the fusion features according to the range of the feature consistency coefficient, and obtains the final prediction result by combining hierarchical decision fusion. Then, it determines the health and emotional state of the sheep according to the preset confidence threshold, divides the early warning level according to the state duration and outputs the corresponding early warning information.

Images