Novel Muscovy duck abnormal behavior automatic identification method

Abstract

The invention relates to the technical field of Muscovy duck breeding, and discloses a novel Muscovy duck abnormal behavior automatic identification method, which comprises the following steps: S1, deploying temperature and humidity, ammonia gas and carbon dioxide concentration sensors at the top of a duck shed according to 5-8m < 2 > / duck shed; 4-8 microphone array sound sensors are arranged in the middle of the peripheral wall at intervals of 3-5 meters; high-definition cameras are arranged on the top diagonally in a crossed manner; each Muscovy duck wears an RFID tag and is associated with a biosensor, and body temperature, heart rate and activity data with the precision of + / -0.1 DEG C are collected at the sampling rate of 100-200 Hz. According to the novel muscovy duck abnormal behavior automatic identification method, multiple sensors are deployed in a duck shed to collect multi-dimensional data, visual data are subjected to feature extraction through an improved convolutional neural network, sound signals are processed through a Mel filter bank, and physiological data calibrate individual differences based on historical data; the multi-modal data are aligned through a dynamic time warping algorithm and then fused through an adaptive weighting model, abnormal behavior characteristics are comprehensively captured, similar behavior modes are effectively distinguished, and the recognition accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of muscovy duck breeding, and specifically to a novel automatic recognition method for abnormal behaviors of muscovy ducks. Background Art

[0002] Abnormal behaviors of muscovy ducks refer to activities that deviate from normal physiological and behavioral patterns during the breeding process, such as feather pecking, abnormal stillness, excessive crowing, etc.; these behaviors not only affect the growth and development of muscovy ducks and egg-laying performance, but may also lead to disease transmission and group harm, causing economic losses to farmers; in a large-scale breeding environment, timely and accurate identification of abnormal behaviors of muscovy ducks is crucial for ensuring breeding efficiency and animal welfare.

[0003] Existing technologies for identifying abnormal behaviors of muscovy ducks mainly rely on single-modal data or simple data processing methods, and have significant defects; on the one hand, traditional methods mostly analyze single-modal data based on vision or environmental sensors, unable to comprehensively capture multi-dimensional features of abnormal behaviors, and difficult to distinguish similar behavior patterns, resulting in low recognition accuracy; on the other hand, in data fusion processing, there is a lack of dynamic weight adjustment and sensor credibility evaluation mechanisms, and it is prone to false alarms and missed detections when facing data noise and sensor fluctuations in a complex breeding environment, unable to meet the requirements of large-scale breeding for accurate identification and real-time early warning of abnormal behaviors, so a novel automatic recognition method for abnormal behaviors of muscovy ducks is proposed. Summary of the Invention

[0004] (1) Technical Problems to be Solved

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a novel automatic recognition method for abnormal behaviors of muscovy ducks, which realizes high-precision recognition and real-time response to abnormal behaviors of muscovy ducks through multi-source data collection, deep feature extraction, intelligent fusion analysis and precise decision-making early warning, effectively improving the intelligent level and economic benefits of breeding, and solving the problems of the existing technology. On the one hand, traditional methods mostly analyze single-modal data based on vision or environmental sensors, unable to comprehensively capture multi-dimensional features of abnormal behaviors, and difficult to distinguish similar behavior patterns, resulting in low recognition accuracy; on the other hand, in data fusion processing, there is a lack of dynamic weight adjustment and sensor credibility evaluation mechanisms, and it is prone to false alarms and missed detections when facing data noise and sensor fluctuations in a complex breeding environment, unable to meet the requirements of large-scale breeding for accurate identification and real-time early warning of abnormal behaviors.

[0006] (2) Technical Solutions

[0007] To achieve the above-mentioned purpose of high-precision identification and real-time response of abnormal behaviors of Muscovy ducks through multi-source data collection, deep feature extraction, intelligent fusion analysis, and precise decision-making and early warning, and effectively improve the intelligent level and economic benefits of breeding, the present invention provides the following technical solutions: A novel automatic identification method for abnormal behaviors of Muscovy ducks, comprising the following steps:

[0008] S1. Deploy temperature and humidity, ammonia, and carbon dioxide concentration sensors on the top of the duck house at intervals of 5 - 8 m 2 / piece; Install 4 - 8 microphone array sound sensors in the middle of the surrounding walls at intervals of 3 - 5 meters; Layout high-definition cameras in a diagonal cross pattern on the top, and collect visual data at a frame rate of 25 - 30 fps and a resolution of 1920×1080; Each Muscovy duck wears an RFID tag, associated with a biosensor, and collects body temperature, heart rate, and activity data with an accuracy of ±0.1 °C at a sampling rate of 100 - 200 Hz;

[0009] S2. Use an improved convolutional neural network to extract visual features. The network contains 4 - 6 convolutional modules, and each module consists of 2 layers of 3×3 convolution (with batch normalization and ReLU activation) cascaded with a 2×2 adaptive pooling layer, and the key region features are strengthened by an attention mechanism at the end; The sound signal is framed with a frame length of 20 - 30 ms, pre-emphasized, and then passed through a 40 - 50-channel Mel filter bank to extract 12 - 16-order Mel frequency cepstral coefficients and short-time zero-crossing rate; Calibrate the individual differences of physiological data based on 7-day historical data;

[0010] S3. Use the dynamic time warping algorithm to align multi-modal data, set a time window of 400 - 600 ms, allowing a deviation of ±80 - 120 ms; Use an adaptive weighted model to fuse features, adjust the modal weights every 3 - 7 minutes according to historical data, and reduce the weights of abnormal fluctuation sensors through sensor credibility evaluation;

[0011] S4. Input the fused features into a cascaded model of Bi-LSTM and SVM. Bi-LSTM extracts 10 - 30-second time series patterns, and SVM classifies with a radial basis kernel function; When the output probability exceeds the dynamic threshold (initially 0.65 - 0.75, adjusted in steps of 0.03 - 0.07 according to the breeding cycle), an early warning is triggered;

[0012] S5. The system generates a report containing the abnormal type, time, location, feature vector summary, and environmental parameter change curve, and sends it to the mobile terminal and the background through the 4G network; Combine the camera coordinates and the 1 - 2㎡ sensor grid to locate the abnormality, and classify the early warning according to the continuous frame number and the cumulative number of times in the area.

[0013] Preferably, the training method of the improved convolutional neural network is as follows:

[0014] Construct 3000 - 5000 pairs of positive and negative samples for normal samples and 1500 - 3000 pairs for abnormal samples, and optimize the feature space using the triplet loss function;

[0015] The initial learning rate is 0.0008 - 0.0012, and it decays to 0.7 - 0.9 times the original every 4 - 6 epochs; The Adam optimizer updates the parameters, with β1 taking 0.85 - 0.95, β2 taking 0.98 - 1.0, and ε taking 1e - 9 - 1e - 7;

[0016] The early stopping method is used to control the training. If the validation set loss does not decrease for 8 - 12 consecutive epochs, the training stops. Image random flipping, rotation, and scaling are used to enhance the generalization ability.

[0017] Preferably, the construction and operation method of the adaptive weighted fusion model is as follows:

[0018] The initial weights of each modality: vision 0.35 - 0.45, physiology 0.25 - 0.35, environment 0.15 - 0.25, sound 0.05 - 0.15; The historical contribution rate database records the modality contributions in real - time and updates the weights every 3 - 7 minutes;

[0019] When the sensor credibility evaluation table detects that the single - sensor fluctuation exceeds 15% - 25% for 3 - 5 consecutive cycles, the corresponding modality weight is reduced by 10% - 30%. A weight threshold is set, and if it is exceeded, manual review is triggered.

[0020] Preferably, the deployment and operation process of the bidirectional long - short - term memory network and support vector machine cascade model is as follows:

[0021] The edge node deploys Bi - LSTM (100 - 300 hidden - layer neurons) to extract temporal patterns, and the data is transmitted to the cloud SVM through a 4G module; The SVM classifies with a radial basis kernel function, and the γ parameter (0.0005 - 150) is optimized by the grid search method;

[0022] Set thresholds according to the breeding cycle: brooding period 0.7 - 0.8, growing period 0.65 - ......

[0023] Preferably, in the process of sound signal processing, pre - emphasis uses a first - order digital filter with parameters 0.93 - 0.99. The center frequencies of the Mel filter bank correspond to the actual frequencies of 280 - 8200Hz, and the logarithmic Mel spectrogram is transformed by the discrete cosine transform to obtain 12 - 16 - order MFCC features.

[0024] Preferably, the physiological data calibration method is as follows: establish a 5-9-day sliding window baseline, and calculate the mean and standard deviation of body temperature and heart rate; when the real-time data deviation exceeds 1.5-2.5 times the standard deviation (body temperature ±0.2-0.4°C, heart rate ±15-25 beats per minute), a warning is triggered; at 1-3 am every day, update the baseline model according to the data of the previous 22-26 hours.

[0025] Preferably, when the dynamic time warping algorithm performs spatio-temporal alignment, construct a similarity matrix of modal data frames with Euclidean distance, set a path planning offset constraint of ±80-120 ms, interpolate the data with inconsistent timestamps, and the synchronization error ≤50-70 ms.

[0026] Preferably, the hierarchical warning mechanism is specifically as follows:

[0027] When the abnormal probability detected in 2-3 consecutive frames exceeds 0.65-0.75, send a first-level warning (yellow), the system generates a brief report including the abnormal type, occurrence location, and brief features, and sends it to the staff's mobile terminal via text message, and at the same time, a pop-up prompt is given in the background management system;

[0028] When 5-8 abnormal triggers are accumulated within 10 minutes in the same grid area, it is upgraded to a second-level warning (red). At this time, a depth report including a detailed analysis of abnormal behaviors, a comparison of environmental data and historical behavior trajectories in the past 20-28 hours, and treatment suggestions is generated, and sent to the staff's mobile terminal in a dual manner of text message and voice call, and an abnormal event special page is generated in the background management system for the staff to comprehensively understand and analyze.

[0029] (III) Beneficial effects

[0030] Compared with the prior art, the present invention provides a new method for automatically identifying abnormal behaviors of muscovy ducks, having the following beneficial effects:

[0031] 1. This new method for automatically identifying abnormal behaviors of muscovy ducks solves the single-modal defect through multi-source data collection and deep feature extraction; a variety of sensors are deployed in the duck house to collect multi-dimensional data, the visual data is extracted with features by an improved convolutional neural network, the sound signal is processed by a Mel filter bank, and the physiological data calibrates individual differences based on historical data; the multi-modal data is aligned by the dynamic time warping algorithm, and then fused by an adaptive weighted model to comprehensively capture the abnormal behavior features, effectively distinguish similar behavior patterns, and improve the recognition accuracy.

[0032] 2. For the automatic recognition method of abnormal behaviors of the new type of muscovy ducks, in data fusion processing, an adaptive weighted fusion model is constructed and a sensor credibility evaluation mechanism is introduced. Each modality has an initial weight. The system records the modality contributions in real time and updates the weights regularly. When the continuous periodic fluctuations of the sensor exceed the limit, the weight of the corresponding modality is reduced, and when the threshold is exceeded, manual review is triggered. This mechanism reduces the influence of data noise and sensor fluctuations, reduces false alarms and missed detections, and meets the requirements of accurate recognition and real-time warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of the automatic recognition method of abnormal behaviors of the new type of muscovy ducks of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the embodiments and drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0035] Please refer to Figure 1 , an automatic recognition method for abnormal behaviors of a new type of muscovy duck, comprising the following steps:

[0036] S1. Deploy temperature and humidity, ammonia, and carbon dioxide concentration sensors on the top of the duck house at 5 - 8 m 2 / piece; install 4 - 8 microphone array sound sensors at intervals of 3 - 5 meters in the middle of the surrounding walls; arrange high-definition cameras in a diagonal cross layout on the top, and collect visual data at a frame rate of 25 - 30 fps and a resolution of 1920×1080; each muscovy duck wears an RFID tag, which is associated with a biosensor, and collects body temperature, heart rate, and activity data with an accuracy of ±0.1 °C at a sampling rate of 100 - 200 Hz;

[0037] S2. Use an improved convolutional neural network to extract visual features. The network contains 4 - 6 convolutional modules. Each module consists of 2 layers of 3×3 convolutions (with batch normalization and ReLU activation) cascaded with a 2×2 adaptive pooling layer. The key region features are strengthened by an attention mechanism at the end; the sound signal is framed with a frame length of 20 - 30 ms, pre-emphasized, and then passed through a 40 - 50-channel Mel filter bank to extract 12 - 16-order Mel frequency cepstral coefficients and short-time zero-crossing rates; calibrate the individual differences of physiological data based on 7-day historical data;

[0038] S3. Use the dynamic time warping algorithm to align multi-modal data, set a time window of 400 - 600 ms, and allow a deviation of ±80 - 120 ms; use an adaptive weighted model to fuse features, adjust the modality weights every 3 - 7 minutes according to historical data, and reduce the weights of abnormal fluctuation sensors through sensor credibility evaluation;

[0039] S4. Input the fused features into the Bi-LSTM and SVM cascade model. The Bi-LSTM extracts the 10-30 second time series pattern, and the SVM classifies it using the radial basis kernel function. When the output probability exceeds the dynamic threshold (initial 0.65-0.75, adjusted in steps of 0.03-0.07 according to the breeding cycle), an alert is triggered.

[0040] S5. The system generates a report containing the anomaly type, time, location, feature vector summary and environmental parameter change curve, and sends it to the mobile terminal and the background through the 4G network; it combines the camera coordinates and the 1-2㎡ sensor grid to locate the anomaly, and issues graded warnings based on the number of consecutive frames and the cumulative number of times in the area.

[0041] Example 1:

[0042] This embodiment describes in detail the specific deployment scheme and operation process of multi-source data collection in a Muscovy duck breeding environment with respect to step S1 in the novel method for automatically identifying abnormal behavior of Muscovy ducks.

[0043] Temperature, humidity, ammonia, and carbon dioxide concentration sensors are evenly deployed on the top of the duck house at a density of 5-8 m2 per sensor. The sensors are connected to the edge computing node via the RS485 bus, and environmental parameters are synchronously collected and CRC verification is performed every 20 seconds.

[0044] A linear array consisting of 4-8 microphones is installed in the middle of the four walls at intervals of 3-5 meters. The microphones are facing the duck activity area and use differential noise reduction technology to suppress environmental noise. Sound signals are collected in real time at a 44.1kHz sampling rate and transmitted to the data preprocessing unit via an RJ45 network cable.

[0045] A 4K high-definition camera is installed at the top diagonal intersection, equipped with a fisheye lens to achieve 180° field of view coverage. It captures video streams at a frame rate of 25-30fps and a resolution of 1920×1080, and transmits them to the edge server via WiFi after H.264 encoding compression.

[0046] Each Muscovy duck wears a UHF band RFID tag with a built-in temperature sensor (accuracy ±0.1°C), a three-axis accelerometer and a heart rate sensor. The tag collects physiological data at a sampling rate of 100-200Hz and sends it to a gateway fixed on the top of the duck house via a BLE low-power Bluetooth module. The gateway packages the data in timestamp order and adds an individual ID.

[0047] All sensors must undergo a 48-hour warm-up calibration before installation to ensure that the temperature and humidity sensor drift is ≤1%RH / ℃, the microphone array phase difference is ≤5°, the camera optical axis deviation is ≤2°, and the RFID tag reading accuracy is ≥99.5%.

[0048] Example 2:

[0049] In this embodiment, for the S2 multi-dimensional feature extraction step in the automatic recognition method of abnormal behaviors of new muscovy ducks, the refined processing flow of visual, sound, and physiological data is described in detail.

[0050] The improved CNN network adopts a 5-layer convolutional module, dynamically adjusts ±1 layer based on the breeding scenario. Each layer constructs a feature extraction unit with 2 groups of 3×3 convolutional kernels, stride 1, padding 1, followed by a batch normalization layer, with γ taken as 1.0, β taken as 0, momentum taken as 0.9, and the LeakyReLU activation function with a negative slope of 0.1. The spatial position information is retained through a 2×2 adaptive pooling layer; the channel attention module is embedded in the 3rd to 5th layers, and the channel weights are generated through global average pooling, and the calculation method is σ(W2·δ(W1·x)), where the dimension of W1 is 256→64, and the dimension of W2 is 64→256, strengthening the spatial feature response of abnormal behaviors such as feather pecking and huddling.

[0051] In the training stage, 42,000 frames of normal behavior samples and 18,000 frames of abnormal samples are used. Normal behaviors include pecking accounting for 28%, resting accounting for 35%, walking accounting for 22%, and preening accounting for 15%. Abnormal samples include feather pecking accounting for 12%, huddling accounting for 25%, abnormal stillness accounting for 38%, and trembling accounting for 25%. Mixed-precision training is adopted, with the precision of FP16, the initial learning rate of 0.001, decaying according to the cosine annealing strategy, the cycle of 100 epochs, the batch size of 64, and the data augmentation methods including random cropping, the scaling ratio from 0.8 to 1.0, and HSV jitter, with the hue of ±5°, the saturation of ±15%, and the brightness of ±10%.

[0052] In the sound signal processing flow, frame segmentation is performed with a frame length of 25 ms, the dynamic adjustment range is ±5 ms, and the frame shift is 10 ms. It passes through a first-order pre-emphasis filter with a coefficient of 0.97±0.02, and then is processed by a 44-channel Mel filter bank with a frequency range of 200 to 8000 Hz and a spacing of 1.0 Mel. 14-order MFCC is extracted, including Δ and ΔΔ coefficients, as well as the short-time zero-crossing rate, with a window of 5 ms, forming a 43-dimensional feature vector; the microphone array is calibrated every shift, that is, every 8 hours. A 1000 Hz sine wave is played with a sound pressure level of 75±2 dB, and the response time difference of each channel is recorded. Channels with a time difference exceeding 50 μs are automatically switched to redundant microphones, and 2 spare microphones are configured for each array.

[0053] Physiological data calibration is based on a 7-day sliding window, including 3 complete day-night cycles. The individual baseline is calculated from 03:00 to 05:00 every day, with the average body temperature μT ± 1.5σT, where σT ≤ 0.3°C, and the average heart rate μH ± 2σH, where σH ≤ 20 bpm. Real-time data is marked as suspected abnormal if it exceeds μT + 1.8σT or μH + 2.5σH, and an alarm is triggered when 3 consecutive sampling points, i.e., 10 seconds, are reached. The temperature of the RFID tag and the environmental sensor are compared every 4 hours. If the deviation > 0.5°C, linear calibration is performed based on the environmental sensor data, with a calibration coefficient ranging from 0.98 to 1.02.

[0054] Example 3:

[0055] This example details the engineering implementation of the dynamic time warping (DTW) algorithm and the adaptive weighted model for the S3 multimodal data fusion step in the automatic recognition method of abnormal behaviors of new Muscovy ducks.

[0056] For spatio-temporal alignment of multimodal data, a time window of 400 to 600 ms with a step size of 50 ms is used to construct a three-dimensional similarity matrix of visual frame levels, sound feature frames, and physiological sampling points.

[0057] The similarity calculation method is as follows: for visual features, cosine distance is used and L2 normalization is performed; for sound features, dynamic time warping distance is used with a window radius of 20 frames; for physiological data, Euclidean distance is used and it is normalized to the [0, 1] interval.

[0058] When path planning, a time deviation of ±100 ms is allowed. The weights for vertical and horizontal movements are set to 1.2, the weight for diagonal movement is set to 1.0, and the endpoint penalty coefficient is 0.8.

[0059] For data with a timestamp deviation > 80 ms, cubic spline interpolation is used with a time resolution of 10 ms. If the interpolation error < 5 ms, it is retained; otherwise, it is marked as invalid data and the sensor self-check process is triggered.

[0060] The adaptive weighted fusion model updates the weights every 5 minutes, allowing a time fluctuation of ±2 minutes. The weight calculation formula is:

[0061]

[0062] Among them, ci is the number of modal contribution times in the past 24 hours, with the visual modality ≥ 3000 times, the physiological modality ≥ 1500 times, the environmental modality ≥ 800 times, and the sound modality ≥ 500 times; acci is the recognition accuracy rate in the past 72 hours, with the visual modality ranging from 0.92 to 0.95, the physiological modality ranging from 0.88 to 0.91, the environmental modality ranging from 0.85 to 0.88, and the sound modality ranging from 0.78 to 0.82; δi is the sensor fluctuation coefficient. When there is a fluctuation > 20% in 4 consecutive cycles, δ is taken as 0.2, and when the fluctuation > 30%, δ is taken as 0.4.

[0063] The fused feature vector is processed by Z - score standardization with a mean of 0 and a standard deviation of 1, and then reduced to 64 dimensions by PCA, retaining 90% of the variance contribution.

[0064] In the sensor credibility evaluation mechanism, if the data fluctuation of the environmental sensor is > 25% in 6 consecutive cycles (i.e., 3 minutes), the corresponding modal weight is reduced by 30% and a yellow - light warning is triggered; if the RFID tag data loss rate is > 5% per hour, the physiological modal weight is reduced to 0.15 and the backup tags are activated. 3 redundant tags are configured for every 50 muscovy ducks.

[0065] The fusion engine has a built - in cache queue with a capacity of 50 groups of data. When the single - modal data interruption is > 20 seconds, it automatically switches to filling with historical data of the same period, and the filling ratio is ≤ 40%.

[0066] Example 4:

[0067] This example details the deployment details of the Bi - LSTM and SVM cascaded model for the S4 dynamic threshold warning step in the new method for automatic recognition of abnormal behaviors of muscovy ducks.

[0068] The Bi - LSTM network is deployed at the edge node, with a bidirectional structure of 128 neurons, stacked in 2 layers, an input time - series window of 15 to 25 seconds, and a step size of 5 seconds, used to capture the time - series patterns of muscovy duck behaviors.

[0069] The activation function uses GELU, the dropout rate is 0.2, and the gradient clipping threshold is 5.0 to prevent gradient explosion.

[0070] The model drift detection is performed every hour. Calculate the KL divergence between the current output and the baseline model at the time of deployment. If > 0.15, online fine - tuning is triggered, using the latest 2000 groups of unlabeled data with a learning rate of 1e - 5.

[0071] Through the model quantization technology, the FP16 precision is converted to INT8, the inference speed is increased by 40%, and the memory occupancy is controlled to < 256MB.

[0072] The cloud - based SVM classifier uses a radial basis kernel function, searches for the optimal parameters through the Bayesian optimization algorithm, with 50 iterations aiming to maximize the F1 - score, and finally determines γ = 0.023 and C = 4.7.

[0073] The classification threshold is dynamically adjusted according to the breeding cycle. The brooding period (i.e., 0 to 4 weeks) has a threshold range of 0.72 to 0.78 with a weekly step size of 0.03; the growing period (i.e., 5 to 12 weeks) has a threshold range of 0.68 to 0.75 with a weekly step size of 0.02; the laying period (i.e., 13 weeks and later) has a threshold range of 0.65 to 0.72 with a weekly step size of 0.015.

[0074] The threshold deviation correction rule is that when it is manually confirmed as normal after 3 consecutive false alarms, the threshold increases by 0.03; when there are 5 consecutive missed alarms, the threshold decreases by 0.05.

[0075] In the early warning response link, the inference delay of the edge node < 80 ms, the classification delay of the cloud < 120 ms, and the confidence level is 95%.

[0076] When the output probability exceeds the threshold, three-level responses are synchronously triggered: First, the key frames of the video are marked in real time, with the range from 3 seconds before the occurrence of the anomaly to 2 seconds after; second, a feature heat map is generated to visualize the contribution degree of each modality to the classification result; third, pre-actions for environmental regulation are started, such as turning on the standby fan 15 seconds in advance.

[0077] The early warning record contains a high-precision timestamp with an accuracy of 10 ms, the position coordinate error is X ± 0.8 m, Y ± 0.6 m, the anomaly type is explained by the SHAP value, and historical similar cases. The top 5 records with a cosine similarity > 0.85 are selected.

[0078] Example 5:

[0079] This example details the location logic of abnormal events, the report generation mechanism, and the multi-level response strategy for the S5 abnormal report and hierarchical early warning steps in the automatic recognition method for abnormal behaviors of new Muscovy ducks.

[0080] When the system detects that the abnormal probability exceeds the threshold, it first performs spatial positioning through the camera coordinates, converts the pixel coordinate system to the physical coordinate system of the duck house, and the positioning accuracy reaches 1 to 2 square meters per grid. At the same time, combined with the environmental sensor grid divided by 5 m × 5 m, the specific location where the anomaly occurs is double-confirmed, and the continuous abnormal frame number and the regional cumulative number are recorded in real time.

[0081] The trigger condition for the first-level early warning (yellow) is that the abnormal probability of 2 consecutive frames is greater than 0.7 or the probability of a single frame is greater than 0.85. At this time, the system generates a brief report, the content of which includes the anomaly type such as pecking and resting timeout, the occurrence time, the specific location, and the 12-dimensional feature vector summary. The breeding personnel are notified synchronously through a text message within 140 characters and a background pop-up window, and the entire response process is controlled within 10 seconds to ensure the preliminary notification and rapid response to abnormal events.

[0082] The triggering conditions for the secondary warning (red) are that there are 6 accumulative anomalies within 10 minutes or the probability of 3 consecutive frames is greater than 0.9 in the same grid area. The system will generate a in-depth analysis report, including the 24-hour environmental parameter change curve, covering key indicators such as temperature, humidity, and ammonia concentration, the heat map of abnormal behavior trajectories, and the comparative analysis of historical similar events. It will be pushed through multiple channels, such as text messages, voice calls, and the background special page. Among them, the voice call adopts a 3-time loop dialing mechanism to ensure that the notification is in place. The report is attached with specific handling suggestions, such as adjusting the ventilation volume and isolating abnormal individuals, providing an operable decision-making basis for breeding personnel.

[0083] The report generation engine adopts a templatized design, supports custom field configuration, and can dynamically adjust display parameters according to different breeding stages, such as the brooding period, growing period, and laying period, to meet diverse monitoring needs.

[0084] The positioning system error calibration mechanism runs once a week. By arranging 10 positioning reference points in the duck house, it collects coordinate data in real time to calculate the deviation matrix between the camera coordinates and the actual position, and then corrects the positioning algorithm parameters to ensure that the long-term positioning accuracy is maintained within 1.5 square meters, guaranteeing the accuracy and reliability of the location information of abnormal events.

[0085] In summary, this new automatic recognition method for abnormal behaviors of Muscovy ducks solves the single-modal defects through multi-source data collection and in-depth feature extraction; a variety of sensors are deployed in the duck house to collect multi-dimensional data. Visual data is used to extract features through an improved convolutional neural network, sound signals are processed by a Mel filter bank, and physiological data is calibrated based on historical data to account for individual differences; multi-modal data is aligned by the dynamic time warping algorithm and then fused using an adaptive weighted model to comprehensively capture abnormal behavior features, effectively distinguish similar behavior patterns, and improve the recognition accuracy.

[0086] Moreover, in the data fusion process of this new automatic recognition method for abnormal behaviors of Muscovy ducks, an adaptive weighted fusion model is constructed and a sensor credibility evaluation mechanism is introduced; each modality has an initial weight, the system records the modality contributions in real time and updates the weights regularly. When the fluctuations of a sensor exceed the limit for consecutive periods, the weight of the corresponding modality is reduced, and when the threshold is exceeded, manual review is triggered; this mechanism reduces the influence of data noise and sensor fluctuations, reduces false alarms and missed detections, meets the requirements of accurate recognition and real-time warning, and solves the problems in the existing technology. On the one hand, traditional methods mostly analyze single-modal data based on vision or environmental sensors, unable to comprehensively capture the multi-dimensional features of abnormal behaviors, difficult to distinguish similar behavior patterns, resulting in low recognition accuracy; on the other hand, in the data fusion process, there is a lack of dynamic weight adjustment and sensor credibility evaluation mechanisms, and it is prone to false alarms and missed detections when facing data noise and sensor fluctuations in a complex breeding environment, unable to meet the requirements of accurate recognition and real-time warning of abnormal behaviors in large-scale breeding.

[0087] All relevant modules involved in this system are hardware system modules or functional modules formed by combining computer software programs or protocols in the prior art with hardware. The computer software programs or protocols themselves involved in such functional modules are all well-known technologies to those skilled in the art and are not the improvements of this system. The improvement of this system lies in the interaction relationships or connection relationships among the modules, that is, the overall structure of the system is improved to solve the corresponding technical problems to be solved by this system.

[0088] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A novel automatic recognition method for abnormal behaviors of muscovy ducks, characterized in that, It includes the following steps: S1. Install temperature and humidity, ammonia, and carbon dioxide concentration sensors on the top of the duck house at a rate of 5 - 8 m 2 / piece; install 4 - 8 microphone array sound sensors in the middle of the surrounding walls at intervals of 3 - 5 meters; deploy high - definition cameras in a diagonal cross - layout on the top to collect visual data at a frame rate of 25 - 30 fps and a resolution of 1920×1080; attach RFID tags to each muscovy duck and associate them with biosensors to collect body temperature, heart rate, and activity data with an accuracy of ±0.1℃ at a sampling rate of 100 - 200 Hz; S2. Use an improved convolutional neural network to extract visual features. The network contains 4-6 convolutional modules. Each module consists of two cascaded 3×3 convolutions (with batch normalization and ReLU activation) followed by a 2×2 adaptive pooling layer. The key region features are enhanced by an attention mechanism at the end. The sound signal is framed with a frame length of 20-30 ms, pre-emphasized, and then passed through a 40-50-channel Mel filter bank to extract 12-16-order Mel frequency cepstral coefficients and short-time zero-crossing rate. Calibrate the individual differences of physiological data based on 7-day historical data; S3. Use the dynamic time warping algorithm to align multi-modal data. Set a time window of 400-600 ms, allowing a deviation of ±80-120 ms. Use an adaptive weighted model to fuse features, and adjust the modal weights every 3-7 minutes according to historical data. Reduce the weights of abnormal fluctuation sensors through sensor credibility evaluation; S4. Input the fused features into a cascaded model of Bi-LSTM and SVM. Bi-LSTM extracts 10-30-second time series patterns, and SVM classifies with a radial basis kernel function; When the output probability exceeds the dynamic threshold (initially 0.65-0.75, adjusted with a step size of 0.03-0.07 according to the breeding cycle), an alarm is triggered; S5. The system generates a report containing the abnormal type, time, location, feature vector summary, and environmental parameter change curve, and sends it to the mobile terminal and the background through the 4G network. Combine the camera coordinates and the 1-2㎡ sensor grid to locate the abnormality, and classify the alarm according to the continuous frame number and the cumulative number of times in the area.

2. The automatic recognition method for abnormal behaviors of a novel muscovy duck according to claim 1, characterized in that The training method of the improved convolutional neural network is as follows: Construct 3000-5000 groups of normal samples and 1500-3000 groups of abnormal samples as positive and negative sample pairs, and optimize the feature space with a triplet loss function; The initial learning rate is 0.0008-0.0012, and it decays to 0.7-0.9 times the original every 4-6 epochs; Update the parameters with the Adam optimizer, where β1 is taken as 0.85-0.95, β2 is taken as 0.98-1.0, and ε is taken as 1e-9-1e-7; Use the early stopping method to control the training. Stop when the validation set loss does not decrease for 8-12 consecutive epochs, and enhance the generalization ability by randomly flipping, rotating, and scaling the images.

3. A novel automatic recognition method for abnormal behaviors of muscovy ducks according to claim 1, characterized in that, The construction and operation method of the adaptive weighted fusion model is as follows: The initial weights of each modality: vision 0.35-0.45, physiology 0.25-0.35, environment 0.15-0.25, sound 0.05-0.15; The historical contribution rate database records the modal contributions in real time and updates the weights every 3-7 minutes; When the sensor credibility evaluation table detects that the fluctuation of a single sensor exceeds 15%-25% for 3-5 consecutive cycles, reduce the corresponding modal weight by 10%-30%, set a weight threshold, and trigger a manual review if it exceeds.

4. A novel automatic recognition method for abnormal behaviors of muscovy ducks according to claim 1, characterized in that, The deployment and operation process of the cascaded model of bidirectional long short-term memory network and support vector machine is as follows: Deploy Bi-LSTM (100-300 hidden layer neurons) at the edge node to extract time series patterns, and transmit the data to the cloud SVM through a 4G module; SVM classifies with a radial basis kernel function, and optimizes the γ parameter (0.0005-150) by the grid search method; Set thresholds according to the breeding cycle: 0.7 - 0.8 for the brooding period, 0.65 - 0.75 for the growing period, and 0.6 - 0.7 for the laying period; when exceeding the threshold, the cloud sends a structured report to the terminal via the MQTT protocol, and the end-cloud delay is 180 - 320 ms.

5. A novel automatic recognition method for abnormal behaviors of muscovy ducks according to claim 1, characterized in that, During the processing of the sound signal, pre-emphasis is performed using a first-order digital filter with parameters of 0.93 - 0.

99. The center frequencies of the Mel filter bank correspond to the actual frequencies of 280 - 8200 Hz. The log Mel spectrum is obtained through discrete cosine transform to get 12 - 16 order MFCC features.

6. The automatic recognition method for abnormal behaviors of a new type of muscovy duck according to claim 1, characterized in that, The physiological data calibration method is as follows: establish a sliding window baseline of 5 - 9 days, calculate the mean and standard deviation of body temperature and heart rate; when the deviation of real-time data exceeds 1.5 - 2.5 times the standard deviation (body temperature ± 0.2 - 0.4 °C, heart rate ± 15 - 25 beats per minute), a warning is triggered; at 1 - 3 o'clock every morning, the baseline model is updated according to the data of the previous 22 - 26 hours.

7. A novel automatic recognition method for abnormal behaviors of muscovy ducks according to claim 1, characterized in that, When the dynamic time warping algorithm performs spatio-temporal alignment, a similarity matrix of modal data frames is constructed with Euclidean distance. The path planning sets an offset constraint of ± 80 - 120 ms, interpolates the data with inconsistent timestamps, and the synchronization error ≤ 50 - 70 ms.

8. A novel automatic recognition method for abnormal behaviors of muscovy ducks according to claim 1, characterized in that, The specific hierarchical warning mechanism is as follows: When the abnormal probability is detected to exceed 0.65 - 0.75 in 2 - 3 consecutive frames, a first-level warning (yellow) is sent. The system generates a brief report containing the abnormal type, occurrence location, and brief features, and sends it to the mobile terminal of the staff via text message. At the same time, a pop-up prompt is given in the background management system. When 5 - 8 abnormal triggers are accumulated within 10 minutes in the same grid area, it is upgraded to a second-level warning (red). At this time, a detailed report containing a detailed analysis of the abnormal behavior, a comparison of the environmental data and historical behavior trajectories in the past 20 - 28 hours, and treatment suggestions is generated, and sent to the mobile terminal of the staff in a dual way of text message and voice call. An abnormal event special page is generated in the background management system for the staff to comprehensively understand and analyze.

Images