Livestock and poultry disease hierarchical filtering adaptive early warning method based on individual exclusive timing reference

Abstract

The application discloses a livestock and poultry epidemic disease hierarchical filtering adaptive early warning method based on individual exclusive timing reference, relates to the technical field of livestock and poultry health artificial intelligence timing data analysis, and comprises the following steps: constructing an exclusive health timing reference library of target livestock and poultry individuals, performing multi-scale timing window feature extraction processing, performing double-layer hierarchical filtering feature discrimination processing, performing daily adaptive reference iterative updating processing, and performing multi-sign coupling hierarchical early warning output processing; the application eliminates the influence of individual physiological differences on early warning accuracy by constructing the exclusive health timing reference of each livestock and poultry individual, adopts a short-time and long-time double-layer filtering cascade architecture, effectively removes daily stress noise such as feeding and fighting, simultaneously extracts multi-dimensional gradient features in combination with a multi-scale timing window, captures slow physiological deviation in a latent period, and cooperates with a daily adaptive reference iteration and multi-factor compensation correction mechanism, so that the early warning threshold can be synchronized with individual growth progress and environmental changes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence time-series data analysis technology for livestock and poultry health, and in particular to an adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks through hierarchical filtering. Background Technology

[0002] In large-scale livestock and poultry farming, identifying the incubation period before disease outbreaks is a crucial step in achieving early intervention and reducing economic losses. Existing livestock and poultry disease early warning technologies mainly employ global fixed threshold early warning schemes or single-layer time-series model early warning schemes.

[0003] Among them, the global fixed threshold early warning scheme sets a uniform threshold for physiological parameters (such as a body temperature exceeding 40.5℃ is considered abnormal) and uses the same judgment standard for all livestock and poultry individuals. However, there are significant physiological differences between different livestock and poultry individuals. Factors such as age, breed, and basal metabolic level all lead to individualized distribution of normal physiological parameters. The global uniform threshold cannot distinguish between normal physiological differences between individuals. In addition, livestock and poultry will experience short-term fluctuations in physiological parameters during daily activities such as feeding, drinking, fighting, and regrouping. This type of scheme also cannot distinguish between such short-term stress fluctuations and the continuous pathological changes caused by diseases, resulting in a large number of false alarms.

[0004] The single-layer time-series model early warning scheme introduces time-series models such as long short-term memory networks to model physiological parameter sequences and attempts to identify abnormal trends from time-series changes. However, this type of scheme only uses a single-layer time-series model to directly model the original signal without separating and processing physiological fluctuations at different time scales. Short-term feeding stress, regrouping stress and the gradual shift of vital signs with an incubation period of several days are superimposed in the time-series signal. The single-layer model is unable to effectively separate the two, resulting in the slow shift of vital signs in the early stage of disease incubation being masked by short-term noise, and thus failing to achieve early warning of the incubation period.

[0005] In summary, existing technologies have several drawbacks, including the neglect of individual differences due to the use of a globally uniform threshold, the inability to separate short-term stress from persistent pathological changes caused by a single-layer time series model, the lack of a multi-scale time series window gradient extraction mechanism making it difficult to capture slow shifts in vital signs during the incubation period of diseases, and the inability to adapt to fluctuations in physiological indicators caused by livestock and poultry growth and seasonal environment due to static and unchanging thresholds.

[0006] Therefore, this invention proposes an adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks through hierarchical filtering to address the problems existing in the prior art. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to propose a hierarchical filtering adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks. This method solves the problems in existing livestock and poultry disease early warning technologies, such as the use of a globally fixed threshold leading to the neglect of individual differences, the inability of single-layer time-series models to distinguish between stress and disease, the lack of individual-differentiated benchmarks, and the inability to adaptively update thresholds.

[0008] To achieve the objectives of this invention, the present invention is implemented through the following technical solution: a hierarchical filtering adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks, comprising the following steps: S1. Construct a dedicated health time-series benchmark database for individual target livestock and poultry. Continuous physiological time-series data of target livestock and poultry individuals under disease-free conditions are collected. Stress anomaly removal processing is performed on the collected continuous physiological time-series data to generate individual-differentiated physiological benchmark intervals for target livestock and poultry individuals. The individual-differentiated physiological benchmark intervals include body temperature benchmark intervals, respiratory rate benchmark intervals, and heart rate benchmark intervals. By establishing a health benchmark for each livestock and poultry individual, the traditional global unified threshold is replaced, eliminating the root cause of false alarms caused by individual physiological differences from the source. S2. Perform multi-scale temporal window feature extraction processing. Three sliding windows with different preset time window widths are set up. Instantaneous feature values ​​in the short-term sliding window, fluctuation amplitude feature values ​​in the medium-term sliding window, and vital sign change gradient feature values ​​in the long-term sliding window are extracted from the real-time physiological time series data of the target livestock and poultry individuals. Multi-vital sign coupling correlation degree calculation is performed on the instantaneous feature values, fluctuation amplitude feature values, and vital sign change gradient feature values ​​to generate a multi-scale time series gradient feature set. The short-term, medium-term, and long-term windows capture instantaneous stress, short-term fluctuations, and long-term gradual trends, respectively. The multi-vital sign correlation degree calculation mines the information on the coordinated changes between body temperature, respiration, and heart rate, providing a feature basis for distinguishing stress from disease for subsequent hierarchical filtering. S3. Perform dual-layer hierarchical filtering feature discrimination processing. Real-time physiological time-series data is input into the first short-time filtering layer. The first short-time filtering layer performs fluctuation pattern recognition processing on the real-time physiological time-series data within a preset time length to generate instantaneous stress fluctuation recognition results. When the instantaneous stress fluctuation recognition results determine that the current time-series fluctuation belongs to instantaneous stress fluctuation, the time-series data segment corresponding to the current time-series fluctuation is filtered out and no warning is triggered. The residual time-series data sequence after processing by the first short-time filtering layer is input into the second long-time filtering layer. The second long-time filtering layer performs vital sign gradient shift recognition processing on the residual time-series data sequence for a duration exceeding a preset duration to generate disease latent feature determination results. The two filtering layers adopt a cascaded architecture. The first layer is specifically responsible for filtering out short-term intermittent stress fluctuations such as feeding, group transfer, and fighting. Only the residual data that passes the screening of the first layer enters the second layer for continuous pathological gradual change analysis. This division of labor mechanism ensures that the warning triggering conditions strictly correspond to disease latent features rather than physiological noise. S4. Perform daily adaptive baseline iterative update processing. The system acquires the health status physiological data of target livestock and poultry individuals at a preset time period. Based on the health status physiological data, it performs fine-tuning and updating of the individual-differentiated physiological benchmark interval to generate the updated individual-differentiated physiological benchmark interval. Then, it performs offset correction processing on the upper and lower limits of the updated individual-differentiated physiological benchmark interval according to the environmental temperature and humidity data and age data to generate adaptive benchmark thresholds. The benchmark interval is fine-tuned daily according to the latest health data of livestock and poultry. At the same time, two external factors, environmental temperature and humidity and age, are introduced to compensate and correct the thresholds, so that the early warning benchmark can be dynamically adjusted with the growth of livestock and poultry and seasonal changes, avoiding missed or false alarms caused by threshold aging. S5. Perform multi-sign coupling hierarchical early warning output processing Based on the results of the disease incubation characteristics assessment, the deviations in body temperature, respiratory rate, and heart rate of the target livestock and poultry individuals are calculated. According to the degree of synchronous deviation of body temperature, respiratory rate, and heart rate deviations, the corresponding level in the preset three-level warning system is matched, and the warning signal associated with the corresponding level is output. The three-level warning system is defined by the degree of synchronous deviation of the three vital signs of body temperature, respiration, and heart rate. The greater the deviation of the vital signs, the higher the warning level, effectively distinguishing between mild and severe disease tendencies and providing farmers with differentiated response decision-making basis.

[0009] Further improvements are made in S1, where continuous physiological time-series data is collected, including body temperature, respiratory rate, and heart rate data for the target livestock and poultry individual over 7 days. The specific steps for stress anomaly removal are as follows: a sliding median filtering algorithm is used to smooth the continuous physiological time-series data, the sliding median value is calculated for each sampling time, and sampling points that deviate from the sliding median value by more than a preset deviation multiple are marked as stress anomalies and removed. The mean and standard deviation of the remaining stable time-series data segment after removing stress anomalies are used as the center value and fluctuation range of the individual-differentiated physiological benchmark interval. The preset deviation multiple is 3 times. A benchmark is established based on 7 days of stable period data. The sliding median filtering effectively removes sudden values ​​caused by short-term stress, ensuring that the established benchmark reflects the true basic physiological state of the livestock and poultry rather than accidental fluctuations.

[0010] Further improvements are made in S2, where the preset time window width of the short-term sliding window is 24 hours, the preset time window width of the medium-term sliding window is 72 hours, and the preset time window width of the long-term sliding window is 7 days. The instantaneous feature value is the measured value of the physiological parameter at the end of the sliding window, the fluctuation amplitude feature value is the difference between the maximum and minimum values ​​of the physiological parameter within the sliding window, and the gradient feature value of the vital sign change is the slope of the linear regression of the physiological parameter from the start to the end of the sliding window. The 24-hour window reflects the real-time state of the day, the 72-hour window captures the fluctuation amplitude change within three days, and the 7-day window quantifies the long-term gradual trend through the linear regression slope. The three scales complement each other to cover the complete time-domain information from transient to steady state.

[0011] A further improvement is made in S2, where the specific steps for performing the multi-sign coupling correlation calculation are as follows: calculate the instantaneous feature value, fluctuation amplitude feature value, and sign change gradient feature value for each of the three sign types: body temperature, respiratory rate, and heart rate. Perform a weighted summation operation on the same type of feature values ​​within the same sliding window for the three sign types according to a preset weight coefficient to generate a multi-sign coupling correlation coefficient for a single time window. By weighted summation, the similar features of the three signs are merged into a single correlation coefficient, which not only preserves the contribution of the independent changes of each sign, but also reflects the degree of coordinated deviation of multiple signs in a comprehensive quantitative form, which is convenient for unified processing by the subsequent filtering layer.

[0012] A further improvement is made in S3, where the first short-time filtering layer adopts a lightweight convolutional neural network structure. This structure contains several one-dimensional convolutional kernels, each with a time coverage of 2 hours. The specific steps for performing fluctuation pattern recognition processing are as follows: the one-dimensional convolutional kernels perform sliding convolution operations on real-time physiological time-series data to extract local time-series fluctuation morphological features within a preset time length. These features are then input into a fully connected classification layer, which outputs a probability value for the instantaneous stress fluctuation category. When the probability value for the instantaneous stress fluctuation category is greater than a preset probability threshold, the current time-series fluctuation is determined to be an instantaneous stress fluctuation. The lightweight convolutional neural network, with its strong ability to extract local morphological features, accurately identifies typical waveform patterns of stress fluctuations such as feeding and fighting within a 2-hour coverage window, quickly filtering out such noise to prevent it from entering the subsequent warning link.

[0013] Further improvements are made in that: the first short-time filtering layer replaces the lightweight convolutional neural network structure with a gated recurrent unit network structure. The gated recurrent unit network structure contains several gated recurrent units, and the time coverage length of the gated recurrent unit is 2 hours. The specific steps for performing fluctuation pattern recognition processing are as follows: the gated recurrent unit performs sequence state update operations on real-time physiological time-series data in chronological order, extracts the temporal state change features within a preset time length, inputs the temporal state change features into the classification output layer, and outputs the instantaneous stress fluctuation category probability value. When the instantaneous stress fluctuation category probability value is greater than the preset probability threshold, it is determined that the current time-series fluctuation belongs to instantaneous stress fluctuation. As an alternative to the lightweight convolutional neural network, the gated recurrent unit's recurrent structure is good at capturing temporal dependencies. Under the condition of maintaining the same 2-hour coverage window, it achieves effective identification of stress fluctuations with lower computational overhead, making it suitable for deployment on edge devices with limited computing resources.

[0014] Further improvements are made in S3, where the second long-term filtering layer adopts a long short-term memory network structure. The long short-term memory network structure contains several long short-term memory units, and the time step coverage of the long short-term memory units is 36 hours. The specific steps for performing the vital sign gradient shift recognition processing are as follows: the residual time-series data sequence is input into the long short-term memory units step by step in chronological order. The long short-term memory units update the hidden state and memory cell state. The hidden state of each time step is input into the temporal attention layer. The temporal attention layer assigns attention weights to the hidden state of each time step, generates a weighted fusion feature vector, and inputs the weighted fusion feature vector into the classification output layer. The classification output layer outputs the probability value of the latent disease feature. When the probability value of the latent disease feature is greater than the preset latent detection threshold, the latent disease feature is detected. The long short-term memory network specifically processes the residual sequence after the first layer of filtering. The 36-hour time step coverage is sufficient to capture the continuous accumulation process of the slow vital sign shift during the latent period of the disease. The introduced attention mechanism further strengthens the focus on key periods, making the model more sensitive to the gradual trend of pathological changes after noise filtering.

[0015] A further improvement is made in S4, where the preset time period is one natural day. The specific steps for fine-tuning and updating the data are as follows: An exponentially weighted moving average is calculated between the mean of the stable time series of the daily health status physiological data and the center value of the current individual-differentiated physiological benchmark interval. The result is used as the center value of the updated individual-differentiated physiological benchmark interval. The weighted average of the standard deviation of the daily health status physiological data and the historical standard deviation is used as the fluctuation range of the updated individual-differentiated physiological benchmark interval. The specific steps for offset correction are as follows: The average ambient temperature, the average relative humidity, and the current age of the target livestock individual are obtained, and the preset ambient temperature compensation coefficient table is queried accordingly. The system includes preset environmental humidity compensation coefficient tables and preset age compensation coefficient tables. It obtains corresponding temperature compensation offset, humidity compensation offset, and age compensation offset. The temperature compensation offset, humidity compensation offset, and age compensation offset are summed to generate a comprehensive compensation offset. The upper and lower limits of the updated individual-differentiated physiological baseline interval are added in the same direction with the comprehensive compensation offset to generate an adaptive baseline threshold. The latest health data is absorbed daily using an exponentially weighted moving average method, giving higher weight to recent data to sensitively track physiological baseline drift. At the same time, the three compensation offsets of environmental temperature, humidity, and age are obtained by looking up tables and superimposed to correct the threshold, ensuring that the baseline's adaptive capability covers the two major variables of growth and season.

[0016] A further improvement is made in S5, where the specific steps for calculating the body temperature offset, respiratory rate offset, and heart rate offset of the target livestock individual are as follows: the current real-time values ​​of the three vital signs (body temperature, respiratory rate, and heart rate) are compared with the upper and lower limits of the corresponding adaptive benchmark thresholds. The offset is zero when the current real-time value of the corresponding vital sign is within the corresponding adaptive benchmark threshold range; the difference between the current real-time value and the upper limit is taken when the current real-time value exceeds the upper limit; and the difference between the current real-time value and the lower limit is taken when the current real-time value is below the lower limit. The offset of each vital sign is measured by the magnitude of its current value deviating from the adaptive benchmark threshold range. Within the range, it indicates normality with no offset. When it exceeds the upper limit or falls below the lower limit, the severity of the deviation is quantified by the distance from the boundary, providing a normalized deviation input value for subsequent graded early warning.

[0017] A further improvement is made in S5, where the specific steps for matching the corresponding level in the preset three-level warning system are as follows: First, determine whether the offset of each of the three types of vital signs is greater than its corresponding preset offset threshold. Then, count the total number of vital sign types whose offsets exceed their respective preset offset thresholds. If the total number of vital sign types is 1, match the first-level warning system; if the total number of vital sign types is 2, match the second-level warning system; and if the total number of vital sign types is 3, match the third-level warning system. The severity of the third-level warning system is higher than that of the second-level warning system, and the severity of the second-level warning system is higher than that of the first-level warning system. The warning level is mapped based on the number of vital signs (temperature, respiration, and heart rate) exceeding their respective offset thresholds. Exceeding one type of vital sign triggers a primary warning to alert the user; exceeding two types of vital signs triggers an intermediate warning suggesting intervention and examination; and simultaneously exceeding all three types of vital signs triggers the highest-level warning prompting emergency response. This tiered mechanism achieves a fine-grained classification of warning intensity.

[0018] The beneficial effects of this invention are as follows: This invention replaces the global fixed threshold with an individualized physiological benchmark, which greatly reduces the basic false alarm rate. On this basis, the overall false alarm rate is further reduced after adding a double-layer hierarchical filter. The individual benchmark provides a normal reference baseline, and the double-layer filter provides an effective change discrimination mechanism. The combination of dual features produces a synergistic noise reduction effect that far exceeds that of single optimization. Moreover, by extracting instantaneous values, fluctuation amplitudes, and change gradients through different time windows, it can capture the gradual shift of signs during the incubation period of diseases that lasts for several days. It can issue early warnings of latent diseases without clinical symptoms, while existing single-layer time series models can only issue alarms after clinical symptoms appear. In addition, by updating the benchmark daily and combining multi-factor compensation correction based on environmental temperature, humidity, and age, the warning threshold can be dynamically adjusted according to the growth of individual livestock and poultry and seasonal environmental changes. It is applicable to various breeding scenarios such as piglets, fattening pigs, and laying hens, without the need for manual recalibration. Attached Figure Description

[0019] Figure 1This is a flowchart illustrating the adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that the technical means not described in detail in the following embodiments are all conventional means in the art, are not the key points of the invention, and will not be elaborated upon.

[0022] Example 1 In this embodiment, 200 healthy pigs were selected as the experimental sample and randomly divided into an experimental group and a control group, with 100 pigs in each group. The experimental group was monitored for disease early warning using the method of this invention, while the control group was monitored for early warning using a traditional global fixed threshold model. Both groups of pigs were raised under the same breeding environment conditions, and various physiological parameters were collected in real time using ear tag-type body temperature sensors, respiratory rate monitoring collars, and heart rate monitoring devices, with a sampling frequency of once every 10 minutes.

[0023] Control group implementation method: The control group used a global fixed threshold early warning scheme. Uniform body temperature thresholds were set at 40.5℃ (upper limit) and 37.5℃ (lower limit), uniform respiratory rate thresholds at 40 breaths / min (upper limit) and 10 breaths / min (lower limit), and uniform heart rate thresholds at 120 beats / min (upper limit) and 60 beats / min (lower limit). An early warning was triggered when any physiological parameter exceeded its corresponding threshold. This scheme uniformly compared all physiological data collected at all time periods using thresholds, without distinguishing between individual differences, short-term stress versus persistent deviations, or performing any filtering.

[0024] See Figure 1 The experimental group implementation method of this embodiment is as follows: S1. Construct a dedicated health time-series benchmark database for individual target livestock and poultry. S1 was performed individually on each pig in the experimental group. Taking pig P-035 as an example, its body temperature, respiratory rate, and heart rate were continuously collected from day 3 to day 9 after entry (144 sampling points per day, for a total of 1008 sampling points). During this period, the pig showed no clinical symptoms, and its feed intake, water intake, and activity were normal. The above time series data were smoothed using a median filtering algorithm with a sliding window width of 11 sampling points. The sliding median value was calculated for each sampling time. Sampling points that deviated from the sliding median value by more than 3 times the standard deviation were marked as stress anomalies and removed (23 stress anomaly sampling points were removed from P-035 during the 7-day collection period, accounting for 2.28% of the total sampling points. The removed data segments mainly corresponded to the short-term rise in body temperature within 30 minutes after feeding). The mean and standard deviation of the remaining stationary time series data segments after removing anomalies were used as the individualized physiological baseline interval for P-035. Calculations showed that the baseline range for body temperature in P-035 pigs was 38.6℃ to 39.8℃ (mean 39.2℃, standard deviation 0.3℃), the baseline range for respiratory rate was 16 breaths / min to 28 breaths / min (mean 22 breaths / min, standard deviation 3 breaths / min), and the baseline range for heart rate was 68 beats / min to 92 beats / min (mean 80 beats / min, standard deviation 8 beats / min). Individualized physiological baseline ranges were constructed for all 100 experimental pigs.

[0025] S2. Perform multi-scale temporal window feature extraction processing. Taking P-035 at a sampling time t (t=14:30) on the 15th day as an example, the instantaneous feature values ​​within the 24-hour window before that time are extracted, namely the measured body temperature of 39.5℃, the measured respiratory rate of 26 breaths / minute, and the measured heart rate of 88 beats / minute at that time; the fluctuation amplitude feature values ​​within the 72-hour window before that time are extracted, namely the difference of 1.4℃ between the maximum body temperature of 39.8℃ and the minimum body temperature of 38.4℃, the difference of 13 breaths / minute between the maximum respiratory rate of 31 breaths / minute and the minimum respiratory rate of 18 breaths / minute, and the difference of 24 breaths / minute between the maximum heart rate of 96 beats / minute and the minimum heart rate of 72 beats / minute; the gradient feature values ​​of vital signs changes within the 7-day window before that time are extracted, namely the slope (unit: ℃ / day), respiratory rate regression slope, and heart rate regression slope obtained by linear regression fitting of all body temperature sampling points within those 7 days. At the current moment, P-035's body temperature gradient is +0.08℃ / day, respiratory rate gradient is +0.7 breaths / minute / day, and heart rate gradient is +1.2 breaths / minute / day. Multi-sign coupling correlation calculation is performed on the features extracted from the above three windows: instantaneous body temperature values, fluctuation amplitudes, and gradients are weighted and summed with weight coefficients of 0.4, 0.3, and 0.3 respectively to obtain the body temperature correlation contribution value; respiratory rate and heart rate are calculated separately in the same way, ultimately generating a multi-scale temporal gradient feature set of P-035 at the current moment.

[0026] S3. Perform dual-layer hierarchical filtering feature discrimination processing. The real-time physiological time-series data of P-035 is input into the first short-time filtering layer. This layer adopts a lightweight convolutional neural network structure, containing 8 one-dimensional convolutional kernels, each with a time coverage of 12 sampling points (corresponding to 2 hours). Each one-dimensional convolutional kernel performs sliding convolution operations on the real-time physiological time-series data to extract local temporal fluctuation morphological features within 2 hours. Taking the body temperature data segment of P-035 from 12:00 to 14:00 on the 15th as an example, during this period, the body temperature showed a peak pattern of rising from 39.2℃ to 39.8℃ and then falling back to 39.3℃. The local temporal fluctuation morphological features extracted by the one-dimensional convolutional kernels are processed by a fully connected classification layer, outputting a transient stress fluctuation category probability value of 0.92 (the preset probability threshold is 0.75). It is determined that the body temperature fluctuation during this period belongs to transient stress fluctuation, and this time-series data segment is filtered out without triggering an alert. According to statistics, P-035 was filtered out during three stress periods on the 15th day (corresponding to the midday feeding period and the afternoon high temperature period respectively), and the filtered data accounted for 7.5% of the total data.

[0027] The residual time-series data sequence processed by the first short-term filtering layer is input into the second long-term filtering layer. This layer adopts a long short-term memory network structure, containing 64 long short-term memory units, with a time step coverage of 216 sampling points (corresponding to 36 hours). Taking the residual body temperature time-series data of P-035 from day 18 to day 20 as an example, the residual sequence shows a continuous and slow upward trend during this period (an average daily increase of about 0.06℃). The long short-term memory units process the residual sequence step by step in chronological order, updating the hidden state and memory cell state. The hidden state of each time step is input into the temporal attention layer. The attention layer assigns attention weights to each time step and generates a weighted fusion feature vector. After processing by the classification output layer, the output disease latency feature probability value is 0.83 (the preset latency judgment threshold is 0.70), indicating that the disease latency feature has been detected.

[0028] S4. Perform daily adaptive baseline iterative update processing. Stable physiological data collected daily from 2:00 AM to 4:00 AM (the resting period for pigs) were used as the daily health status physiological data. Taking the update of P-035 on the 15th day as an example: the average body temperature during the stable period was 39.3℃, and an exponentially weighted moving average (smoothing factor α=0.2) was performed with the current baseline center value of 39.2℃, resulting in an updated baseline center value of 39.22℃; the standard deviation of body temperature during the stable period was 0.28℃, and the weighted average (with each weight of 0.5) of the standard deviation of 0.30℃ was used as the updated fluctuation range of 0.29℃. The average ambient temperature of day 15 is 26.5℃, the average relative humidity is 72%, and the current age of P-035 is 85 days. The preset ambient temperature compensation coefficient table (26.5℃ corresponds to temperature compensation offset +0.08℃), the preset ambient humidity compensation coefficient table (72% corresponds to humidity compensation offset +0.03℃), and the preset age compensation coefficient table (85 days corresponds to age compensation offset +0.05℃) are queried. The sum of the three is used to generate a comprehensive compensation offset of +0.16℃. The updated upper and lower limits of the body temperature benchmark interval (38.93℃~39.51℃) are added in the same direction with the comprehensive compensation offset to generate the upper limit of the adaptive benchmark threshold of 39.67℃ and the lower limit of 38.77℃ for the day.

[0029] S5. Perform multi-sign coupling hierarchical early warning output processing At 14:30 on day 21, P-035 tested positive for the incubation characteristics of the disease as determined by S3. The offsets of the current real-time values ​​of the three vital signs at this moment from the corresponding adaptive baseline thresholds (after daily updates in step S4, the baseline range for P-035 on day 21 is 38.8℃~39.6℃ for body temperature, 17~29 breaths / minute for respiratory rate, and 69~93 beats / minute for heart rate) were calculated. At this moment, P-035's body temperature was 40.1℃, exceeding the upper limit of 39.6℃ by an offset of 0.5℃; respiratory rate was 34 breaths / minute, exceeding the upper limit of 29 breaths / minute by an offset of 5 breaths / minute; and heart rate was 96 beats / minute, exceeding the upper limit of 93 beats / minute by an offset of 3 beats / minute. The preset offset thresholds for the three types of vital signs are 0.3℃ for body temperature, 4 breaths / minute for respiratory rate, and 2 beats / minute for heart rate. The offset of each of the three types of vital signs is greater than its respective preset offset threshold. The total number of vital sign types is 3, which matches the three-level warning level and outputs a three-level warning signal.

[0030] Experimental results: In a 60-day monitoring trial, 8 out of 100 pigs in the control group actually developed the disease. The global fixed threshold scheme missed all 8 diseased samples (i.e., no warning was given before the appearance of clinical symptoms), triggering alarms only after the clinical symptoms became obvious, with an average of 42 false alarms per day. Among these, 6 pigs did not develop any disease during the entire 60-day trial period, but were falsely alarmed multiple times by the global threshold scheme due to factors such as a short-term increase in body temperature after feeding and an increased heart rate caused by fighting. In the experimental group, 9 out of 100 pigs actually developed the disease (P-035 was the first confirmed case on day 21). The method of this embodiment successfully detected 8 out of the 9 diseased pigs, missing only one individual with extremely rapid disease progression, with an average of only 6.8 false alarms per day and an average early warning time of 4.5 days before the onset of the disease. The main cause of false alarms was transient data jumps caused by poor sensor contact. Statistical results from the control and experimental groups show that the method of this invention is significantly superior to the existing global fixed threshold scheme in terms of false alarm rate, false alarm rate, and early warning time.

[0031] Comparative Analysis: Validation of the Synergistic Effect of Individual Baseline and Two-Layer Filtering To verify the synergistic effect of combining individual-specific benchmarks and two-layer hierarchical filtering in this embodiment, a comparative scheme of "individual benchmarks only + no two-layer filtering" was set up under the same experimental environment and the same sample size (100 pigs). This scheme only performs S1 to construct individual-differentiated physiological benchmark intervals, and then directly compares real-time physiological data with the benchmark threshold to trigger an early warning. It does not perform multi-scale feature extraction in S2, two-layer hierarchical filtering in S3, or adaptive update in S4, and directly uses the static benchmark threshold constructed in S1 for early warning judgment. The comparative results show that when only individual-specific benchmarks are used without two-layer hierarchical filtering, the average number of false alarms per day is 17.9, and the early warning time is only 0.5 days before the onset of disease. Although individual benchmarks eliminate some false alarms caused by individual differences, they cannot distinguish between short-term stress fluctuations and continuous pathological changes. Short-term fluctuations in vital signs caused by daily activities such as feeding and fighting still frequently trigger false alarms. At the same time, due to the lack of multi-scale gradient feature extraction, this scheme cannot capture slow shifts in vital signs during the incubation period and can only issue an early warning after the vital sign value exceeds the threshold. This demonstrates that only by combining individual-specific benchmarks and two-layer hierarchical filtering can excellent results be achieved, and there is a clear synergistic effect between the two.

[0032] Example 2 To further verify the universality of this invention across different livestock and poultry species, this embodiment selected 500 laying hens from a laying hen farm for supplementary verification experiments. Considering the physiological characteristics of poultry (normal body temperature range 40.5℃~42.0℃, respiratory rate 15~30 breaths / minute, heart rate 250~450 beats / minute), the warning parameters of the method of this invention were adaptively adjusted as follows: In S1, continuous physiological time-series data of each laying hen under disease-free conditions were collected for 5 consecutive days (due to the high metabolic rate and short baseline stability period of poultry, the number of days for collection was adjusted from 7 days to 5 days). Taking one laying hen with the number L-078 as an example, its baseline range of body temperature was 41.0℃~41.8℃ (mean 41.4℃, standard deviation 0.2℃), the baseline range of respiratory rate was 18~26 breaths / minute (mean 22 breaths / minute, standard deviation 3 breaths / minute), and the baseline range of heart rate was 280~360 beats / minute (mean 320 beats / minute, standard deviation 25 beats / minute). In S2, the parameters of the three-layer sliding window are adjusted accordingly to a short window of 12 hours (poultry have a high metabolic rate and a shorter stress response cycle), a medium window of 48 hours, and a long window of 5 days. In S3, the convolution kernel time coverage length of the first short-time filter layer is adjusted to 1 hour, and the long short-term memory unit time step coverage length of the second long-time filter layer is adjusted to 24 hours. In S4, the daily adaptive iteration update frequency remains unchanged (still 1 calendar day). The environmental temperature and humidity compensation coefficient table and the age compensation coefficient table are recalibrated according to the physiological characteristics of laying hens (the body temperature baseline of laying hens during the laying period is about 0.2℃ higher than that during the non-laying period, and the corresponding age compensation offset is set separately). The grading rules for the three-level early warning system in S5 remain unchanged, and the preset offset thresholds for each vital sign are reset according to the normal fluctuation range of poultry. After 90 days of monitoring and verification, the average advance warning time for the incubation period of avian influenza in laying hens using the method of this invention is 3.2 days, with an average daily false alarm rate of 5.9%, verifying the universality of the method of this invention across different livestock and poultry species and the flexibility of parameter adaptation.

[0033] In this embodiment, the lightweight convolutional neural network, long short-term memory network, and gated recurrent unit network structures all require pre-training before use. The training data comes from historical aquaculture data, specifically physiological time-series data labeled with stress event periods and confirmed case incubation periods. For the first short-term filtering layer (lightweight convolutional neural network or gated recurrent unit), the training labels are binary classification labels of "instantaneous stress fluctuations" and "non-instantaneous stress fluctuations," with the training objective being to minimize the classification cross-entropy loss function. For the second long-term filtering layer (long short-term memory network), the training labels are binary classification labels of "disease incubation period" and "normal period," with the training objective being to minimize the classification cross-entropy loss function. The weight parameters of the attention layer are automatically learned during end-to-end training and do not require separate training. The specific training methods for each network structure are common knowledge in the field and will not be elaborated upon in this invention.

[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hierarchical filtering adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks, characterized in that, Includes the following steps: S1. Collect continuous physiological time-series data of target livestock and poultry individuals under disease-free conditions and perform stress abnormal value removal processing to generate individual-differentiated physiological baseline intervals for target livestock and poultry individuals, including body temperature baseline interval, respiratory rate baseline interval and heart rate baseline interval. S2. Set sliding windows with different preset time window widths, extract instantaneous feature values ​​in the short-term sliding window, fluctuation amplitude feature values ​​in the medium-term sliding window, and physical sign change gradient feature values ​​in the long-term sliding window from the real-time physiological time series data of the target livestock and poultry individuals, and perform multi-physical sign coupling correlation degree calculation processing to generate a multi-scale time series gradient feature set. S3. Input the real-time physiological time series data into the first short-time filtering layer, perform fluctuation pattern recognition processing within a preset time length, generate instantaneous stress fluctuation recognition results, when the instantaneous stress fluctuation recognition results determine that the current time series fluctuation belongs to instantaneous stress fluctuation, filter out the time series data segment corresponding to the current time series fluctuation, do not trigger an alarm, input the residual time series data sequence after processing by the first short-time filtering layer into the second long-time filtering layer, perform vital sign gradient shift recognition processing on the residual time series data sequence for a duration exceeding a preset duration, and generate disease latent feature determination results; S4. Obtain the physiological data of the health status of the target livestock and poultry individuals at a preset time period, perform fine-tuning and updating processing on the individual-differentiated physiological benchmark interval, generate the updated individual-differentiated physiological benchmark interval, and perform offset correction processing on the upper and lower limit thresholds of the updated individual-differentiated physiological benchmark interval based on the environmental temperature and humidity data and age data to generate an adaptive benchmark threshold. S5. Based on the results of the disease incubation characteristics determination, calculate the body temperature deviation, respiratory rate deviation, and heart rate deviation of the target livestock and poultry individual. Based on the degree of synchronous deviation of the body temperature deviation, respiratory rate deviation, and heart rate deviation, match the corresponding level in the preset three-level warning level and output the warning signal associated with the corresponding level.

2. The adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to claim 1, characterized in that: In step S1, when collecting continuous physiological time-series data, the body temperature time-series data, respiratory rate time-series data, and heart rate time-series data of the target livestock and poultry individuals are continuously collected over 7 days. The specific steps for performing stress anomaly removal processing are as follows: the continuous physiological time-series data are smoothed using a moving median filtering algorithm, the moving median value at each sampling time is calculated, and sampling points that deviate from the moving median value by more than a preset deviation multiple are marked as stress anomalies and removed. The mean and standard deviation of the remaining stable time-series data segment after removing stress anomalies are used as the center value and fluctuation range of the individual differentiated physiological benchmark interval. The preset deviation multiple is 3 times.

3. The adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to claim 1, characterized in that: In S2, the preset time window widths of the short-term sliding window, the medium-term sliding window, and the long-term sliding window are 24 hours, 72 hours, and 7 days, respectively. The instantaneous feature value is the measured value of the physiological parameter at the end of the sliding window. The fluctuation amplitude feature value is the difference between the maximum and minimum values ​​of the physiological parameter within the sliding window. The gradient feature value of the vital sign change is the slope of the linear regression of the physiological parameter from the start time to the end time of the sliding window.

4. The adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to claim 1, characterized in that: In step S2, the specific steps for performing the multi-sign coupling correlation degree calculation are as follows: calculate the instantaneous feature value, fluctuation amplitude feature value, and sign change gradient feature value for each of the three sign types: body temperature, respiratory rate, and heart rate. Perform a weighted summation operation on the same type of feature value within the same sliding window of the three sign types according to the preset weight coefficient to generate the multi-sign coupling correlation degree coefficient of a single time window.

5. The adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to claim 1, characterized in that: In S3, the first short-time filtering layer adopts a lightweight convolutional neural network structure. The lightweight convolutional neural network structure contains several one-dimensional convolutional kernels. The time coverage length of the one-dimensional convolutional kernels is 2 hours. The specific steps for performing fluctuation pattern recognition processing are as follows: the one-dimensional convolutional kernels perform sliding convolution operations on real-time physiological time-series data to extract local time-series fluctuation morphological features within a preset time length. The local time-series fluctuation morphological features are input into a fully connected classification layer, and the instantaneous stress fluctuation category probability value is output. When the instantaneous stress fluctuation category probability value is greater than a preset probability threshold, the current time-series fluctuation is determined to be an instantaneous stress fluctuation.

6. The adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to claim 5, characterized in that: The first short-time filtering layer replaces the lightweight convolutional neural network structure with a gated recurrent unit network structure. The gated recurrent unit network structure contains several gated recurrent units, and the time coverage length of the gated recurrent units is 2 hours. The specific steps for performing fluctuation pattern recognition processing are as follows: the gated recurrent unit performs sequence state update operations on real-time physiological time-series data in chronological order, extracts the time-series state change features within a preset time length, inputs the time-series state change features into the classification output layer, and outputs the instantaneous stress fluctuation category probability value. When the instantaneous stress fluctuation category probability value is greater than the preset probability threshold, it is determined that the current time-series fluctuation belongs to the instantaneous stress fluctuation.

7. The adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to claim 1, characterized in that: In S3, the second long-term filtering layer adopts a long short-term memory network structure. The long short-term memory network structure contains several long short-term memory units. The time step coverage length of the long short-term memory unit is 36 hours. The specific steps for performing the vital sign gradient offset recognition processing are as follows: the residual time-series data sequence is input into the long short-term memory unit step by step in chronological order. The long short-term memory unit updates the hidden state and memory cell state. The hidden state of each time step is input into the temporal attention layer. The temporal attention layer assigns attention weights to the hidden state of each time step to generate a weighted fusion feature vector. The weighted fusion feature vector is input into the classification output layer. The classification output layer outputs the probability value of the latent disease feature. When the probability value of the latent disease feature is greater than the preset latent judgment threshold, the latent disease feature is detected.

8. The adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to claim 1, characterized in that: In step S4, the preset time period is one natural day. The specific steps for fine-tuning and updating the data are as follows: The mean of the stable time series of the daily health status physiological data is used to perform an exponentially weighted moving average calculation with the center value of the current individual-differentiated physiological benchmark interval. The result is used as the center value of the updated individual-differentiated physiological benchmark interval. The weighted average of the standard deviation of the daily health status physiological data and the historical standard deviation is used as the fluctuation range of the updated individual-differentiated physiological benchmark interval. The specific steps for offset correction are as follows: The average ambient temperature, the average relative humidity, and the current age of the target livestock individual are obtained. The preset ambient temperature compensation coefficient table, the preset ambient humidity compensation coefficient table, and the preset age compensation coefficient table are queried respectively to obtain the corresponding temperature compensation offset, humidity compensation offset, and age compensation offset. The temperature compensation offset, humidity compensation offset, and age compensation offset are summed to generate a comprehensive compensation offset. The upper and lower limit thresholds of the updated individual-differentiated physiological benchmark interval are added in the same direction with the comprehensive compensation offset to generate an adaptive benchmark threshold.

9. The adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to claim 1, characterized in that: In step S5, the specific steps for calculating the body temperature offset, respiratory rate offset, and heart rate offset of the target livestock individual are as follows: the current real-time values ​​of the three vital signs of body temperature, respiratory rate, and heart rate are compared with the upper and lower limits of the corresponding adaptive benchmark thresholds, respectively. The offset is zero when the current real-time value of the corresponding vital sign is within the range of the corresponding adaptive benchmark threshold, the difference between the current real-time value and the upper limit is taken when the current real-time value exceeds the upper limit, and the difference between the current real-time value and the lower limit is taken when the current real-time value is lower than the lower limit.

10. The adaptive early warning method for livestock and poultry diseases based on individual-specific time-series benchmarks according to claim 1, characterized in that: In step S5, the specific steps for matching the corresponding level in the preset three-level warning level are as follows: determine whether the offset of each of the three types of vital signs is greater than their respective preset offset thresholds, count the total number of vital sign types whose offsets are greater than their respective preset offset thresholds, match the first-level warning level when the total number of vital sign types is 1, match the second-level warning level when the total number of vital sign types is 2, and match the third-level warning level when the total number of vital sign types is 3. The severity of the third-level warning level is higher than that of the second-level warning level, and the severity of the second-level warning level is higher than that of the first-level warning level.

Images