Fresh agricultural product transportation quality monitoring system for in-transit logistics

Abstract

The invention discloses a fresh agricultural product transportation quality monitoring system for in-transit logistics, and relates to the technical field of transportation monitoring. The fresh agricultural product transportation quality monitoring system comprises a multi-path signal acquisition unit, an edge calculation unit, a data processing module, a quality scoring module and a route selection module, the multi-path signal acquisition module integrated sensor comprises a temperature probe, a capacitive humidity sensor, a photoresistor and a high-resolution RGB camera in a carriage, and the sensor is responsible for acquiring the temperature T, humidity RH, illumination Lux and RGB images in the carriage and transmitting the images to the edge calculation module; according to the invention, an ARI MA-LSTM hybrid model is constructed in the quality scoring module, processing and modeling analysis are carried out by using historical temperature, humidity, illumination, chromatic aberration and real-time quality score multi-source information, high-precision prediction of the quality of agricultural products in the future 6 hours is realized, and a decision basis is provided for route selection based on dynamic comparison between a prediction result and a preset threshold value, so that the quality of the agricultural products can be accurately predicted. And the perspectiveness and reliability of quality prediction are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of transportation monitoring, and in particular to a fresh agricultural product transportation quality monitoring system for in-transit logistics. Background Art

[0002] In modern cold chain logistics, the transportation of fresh agricultural products faces multiple challenges. First, the transportation process spans multiple environments—open air, warehouses, and inside vehicles—resulting in frequent changes in temperature, humidity, light, and other conditions. Traditional constant temperature boxes and refrigeration equipment, while able to maintain a low temperature environment at a macro level, are unable to conduct real-time, refined, multi-point monitoring of the microenvironment inside the box. The quality of agricultural products themselves will deteriorate over time, manifesting as water loss, color changes, and even bacterial proliferation. A single temperature or humidity alarm can only reflect the instantaneous state, but cannot capture the cumulative effects of quality degradation and different environmental factors. For example, a sudden temperature rise accelerates water evaporation, while a high humidity environment can lead to rapid microbial growth. The combined effect of these two factors can cause quality to decline much faster than the sum of any single factor. Modern supply chains require a close integration of quality monitoring and logistics scheduling. When quality risks arise, not only must an immediate alert be issued, but the shortest route, real-time road conditions, and delivery timeliness must be comprehensively considered to flexibly adjust itineraries or activate backup preservation measures. Existing systems, lacking multi-source data fusion and time series prediction capabilities, struggle to effectively link quality warnings with route optimization, thus failing to provide forward-looking support for delivery decisions.

[0003] Existing systems often rely on simple threshold alarms, making static judgments on a single environmental parameter and ignoring the coupling relationships between multidimensional factors such as temperature, humidity, light, and product color. This makes it difficult to effectively capture the nonlinear fluctuations and time lag effects of agricultural product quality during logistics.

[0004] In view of the above technical defects, a solution is now proposed. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the present invention provides a fresh agricultural product transportation quality monitoring system for in-transit logistics.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A fresh agricultural product transportation quality monitoring system for in-transit logistics, comprising: a multi-channel signal acquisition unit, an edge computing unit, a data processing module, a quality scoring module, and a route selection module;

[0007] The multi-channel signal acquisition module integrates sensors including a temperature probe, a capacitive humidity sensor, a photoresistor, and a high-resolution RGB camera inside the vehicle. The sensors are responsible for collecting the internal temperature T, humidity RH, and light Lux, and transmitting the RGB image to the edge computing module.

[0008] The edge computing module pre-processes temperature T, humidity RH, light Lux and RGB images;

[0009] The data processing module performs linear compensation calibration for temperature T and humidity RH; performs color conversion on the RGB image and selects a representative area, and calculates the color difference ΔE of the agricultural product surface ten minutes before and after in the representative area;

[0010] The quality scoring module is responsible for weighted comprehensive evaluation and outputs the quality index QI. It uses temperature T, humidity RH, light Lux, color difference ΔE, and the quality index QI to build an ARIMA-LSTM hybrid model. The ARIMA-LSTM hybrid model predicts the quality prediction sequence within the next 6 hours.

[0011] The route selection module monitors the quality index QI of the arrival time in the quality prediction sequence. When the quality index QI of the monitored arrival time falls below the preset threshold Q_warn, the delivery driver is reminded based on the delivery time.

[0012] The edge computing module performs raw data anomaly detection and noise suppression;

[0013] The raw data anomaly detection includes:

[0014] S201. Compare each sampling value with the upper and lower limits of the range. If the temperature T is not within the range of -55°C to 125°C, the humidity RH is not within the range of 0%RH to 100%RH, or the illumination Lux is not within the range of 0.01Lux to 100,000Lux, mark the sampling value as abnormal and set it to missing.

[0015] S202, calculate the difference ΔT, ΔRH, and ΔLux between two adjacent samplings. When |ΔT|>5°C, |ΔRH|>15%RH, or |ΔLux|>2000Lux, treat the corresponding sampling value as a transient jump and fill it with the previous and next valid sampling values ​​in a linear interpolation manner, and record the jump event at the same time;

[0016] S203. When three consecutive samples are abnormal or missing, backfill according to the ratio of 70%:30% based on the most recent valid value and the average value of the historical window.

[0017] The noise suppression includes:

[0018] S211 , performing a double exponential smoothing algorithm on the temperature T, recursively updating the level and trend components according to the preset smoothing coefficient α_T = 0.2 and trend coefficient β_T = 0.1;

[0019] S212, performing one-dimensional Kalman filtering and adaptive low-pass filtering on the humidity RH in sequence, the former updating the optimal estimate in real time based on the state space model, and the latter dynamically adjusting the cutoff frequency according to the variance of the observation sequence;

[0020] S213 , performing N=7 median filtering to eliminate extreme values ​​and three-point simple averaging processing on the illumination Lux in sequence.

[0021] The data processing module performs linear compensation calibration on temperature T and humidity RH, including:

[0022] S301, based on the difference ΔT between the initial temperature value and the reference temperature of 25°C, the humidity is corrected by the coupling coefficient α to obtain the corrected humidity RH2, the formula is RH2 = RH1 – α·ΔT, and RH2 is limited to be maintained in the range of 0 to 100%;

[0023] S302: Using the difference ΔRH between the corrected humidity RH2 and the reference humidity of 50% RH, combined with the humidity-temperature coupling coefficient β, a secondary correction is performed on the initial temperature value to obtain a final temperature compensation value T2. The calculation method is T2 = T1 – β·ΔRH, and T2 is limited to a safe range of -55°C to 125°C.

[0024] S303 , adopting an iterative convergence strategy, taking T1 and RH1 as initial iteration values, and alternately performing the above temperature compensation and humidity compensation steps three times, and updating the temperature and humidity calibration results in each iteration.

[0025] The data processing performs color conversion on the RGB image and selects a representative area, including:

[0026] S401, converting the RGB image from the RGB color space to the HSV color space, and performing pixel-level filtering based on a preset H, S, and V threshold range to generate a preliminary segmentation mask Mask1;

[0027] S402, applying morphological opening and closing operations to Mask1 to remove isolated noise points and fill small holes;

[0028] S403, performing connected domain analysis on Mask2, sorting the connected domain pixels from large to small according to their area, and selecting the connected domain region with the largest area as the representative region Patch_i;

[0029] S404 , convert the representative region Patch_i and the corresponding representative region Patch_i_ref ten minutes ago into Lab space respectively, and obtain (Li, ai, bi) and (Li_ref, ai_ref, bi_ref);

[0030] S405. Calculate ΔE_pixel on all pixels in the representative area Patch_i. The calculation formula is as follows:

[0031]

[0032] S406 , averaging all ΔE_pixels in the representative region Patch_i to obtain the color difference ΔE of the representative region Patch_i.

[0033] The quality scoring module performs range mapping and normalization on temperature T, humidity RH, light Lux, and color difference ΔE to between 0 and 1; and performs weighted summation on each feature according to a preset weight vector to obtain the original quality score S_raw.

[0034] The quality scoring module constructs an ARIMA-LSTM hybrid model including:

[0035] S501, arranging temperature T, humidity RH, light Lux, color difference ΔE, and original mass score S_raw in time series to obtain an original feature sequence;

[0036] S502, performing difference preprocessing on the original feature sequence to obtain the Sraw sequence to remove non-stationary components, and the difference order is determined by the unit root test ADF;

[0037] S503, using the autocorrelation function ACF and the partial autocorrelation function PACF of the Sraw sequence to determine the ARIMA order (p, d, q);

[0038] S504, applying the maximum likelihood estimation method to fit the ARIMA model to obtain a linear prediction function;

[0039] S505: Repeat the above steps to sequentially predict the quality index sequence within the next 6 hours.

[0040] The route selection module receives a six-hour quality prediction sequence and obtains real-time traffic information for each candidate delivery route from a navigation map tool, including estimated travel time, road congestion factor, and road section safety factor. Based on the arrival time of each route, the module extracts the quality index corresponding to that arrival time from the prediction sequence. The module compares each route's arrival quality index with a preset warning threshold, marks routes below the threshold as risky routes, and sends these risky routes to the driver.

[0041] The present invention provides a fresh agricultural product transportation quality monitoring system for in-transit logistics. Compared with the existing technology, it has the following advantages:

[0042] The present invention constructs an ARIMA-LSTM hybrid model in the quality scoring module and uses multi-source information such as historical temperature, humidity, light, color difference, and real-time quality score for processing and modeling analysis to achieve high-precision prediction of agricultural product quality within the next 6 hours. Based on the dynamic comparison of the prediction results with preset thresholds, it provides a decision-making basis for route selection, effectively improving the foresight and reliability of quality prediction.

[0043] The present invention integrates a multi-channel signal acquisition and edge pre-processing unit for temperature, humidity, light, and high-resolution RGB images to achieve multi-dimensional real-time monitoring and anomaly verification of the carriage environment and the surface status of agricultural products. It can eliminate noise and anomalies at the source of the data, ensuring subsequent color difference calculation and quality assessment, and significantly improving the accuracy and stability of monitoring data.

[0044] The present invention monitors the quality index and warning threshold of the arrival time in the predicted quality sequence in real time, and reminds the driver in combination with the delivery time. It can issue a warning in time before the quality of agricultural products falls below the safety line, guide the vehicle to adjust the itinerary or take preservation measures, thereby minimizing quality losses and logistics costs during transportation, and improving overall delivery efficiency and product safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the principle framework of the present invention. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] See also Figure 1 , the present application provides a fresh agricultural product transportation quality monitoring system for in-transit logistics, including: a multi-channel signal acquisition unit, an edge computing unit, a data processing module, a quality scoring module, and a route selection module;

[0048] The multi-channel signal acquisition module integrates sensors including a temperature probe, a capacitive humidity sensor, a photoresistor, and a high-resolution RGB camera inside the vehicle. The sensors are responsible for collecting the internal temperature T, humidity RH, and light Lux, and transmitting the RGB image to the edge computing module.

[0049] The edge computing module pre-processes temperature T, humidity RH, light Lux and RGB images;

[0050] The data processing module performs linear compensation calibration for temperature T and humidity RH; performs color conversion on the RGB image and selects a representative area, and calculates the color difference ΔE of the agricultural product surface ten minutes before and after in the representative area;

[0051] The quality scoring module is responsible for weighted comprehensive evaluation and outputs the quality index QI. It uses temperature T, humidity RH, light Lux, color difference ΔE, and the quality index QI to build an ARIMA-LSTM hybrid model. The ARIMA-LSTM hybrid model predicts the quality prediction sequence within the next 6 hours.

[0052] The system hardware includes: Temperature probe: using the DS18B20 digital temperature sensor, with a temperature measurement range of -55 to +125°C and a resolution of 0.1°C. It is mounted via a 1-Wire bus and supports multi-point cascade layout; each probe has its own unique serial number to achieve online identification and positioning.

[0053] Humidity sensor: Utilizes a high-precision capacitive humidity module (range 0 to 100% RH, accuracy ±1% RH), outputs an analog voltage, sampled by an on-chip 16-bit Σ-Δ ADC with a configurable sampling frequency (1 to 20 Hz); includes built-in temperature compensation and calibration coefficient storage.

[0054] Light sensor: A silicon photoresistor is used in conjunction with a constant current excitation circuit to achieve a wide range of 0.01 to 100,000 Lux. Data is collected through a 12-bit ADC and supports manual / automatic gain switching (AGC). The module also has an integrated light shield to suppress direct interference.

[0055] Image acquisition: The vehicle is equipped with a cascaded dual-camera solution in the cabin. The main camera is a 2MP RGB CMOS camera (1920×1080@30fps), and the auxiliary camera is an 850nm infrared night vision camera. Both cameras support hardware-level distortion correction, electronic image stabilization (EIS), automatic exposure (AE), and automatic white balance (AWB).

[0056] The route selection module monitors the quality index QI of the arrival time in the quality prediction sequence. When the quality index QI of the monitored arrival time falls below the preset threshold Q_warn, the delivery driver is reminded based on the delivery time.

[0057] The edge computing module performs anomaly detection and noise suppression on raw data;

[0058] Raw data anomaly detection includes:

[0059] S201. Compare each sampling value with the upper and lower limits of the range. If the temperature T is not within the range of -55°C to 125°C, the humidity RH is not within the range of 0%RH to 100%RH, or the illumination Lux is not within the range of 0.01Lux to 100,000Lux, mark the sampling value as abnormal and set it to missing.

[0060] S202, calculate the difference ΔT, ΔRH, and ΔLux between two adjacent samplings. When |ΔT|>5°C, |ΔRH|>15%RH, or |ΔLux|>2000Lux, treat the corresponding sampling value as a transient jump and fill it with the previous and next valid sampling values ​​in a linear interpolation manner, and record the jump event at the same time;

[0061] S203. When three consecutive samples are abnormal or missing, backfill according to the ratio of 70%:30% based on the most recent valid value and the average value of the historical window.

[0062] In the edge computing module, raw data anomaly detection first compares each sampled value with the preset upper and lower limits. Sampling points that exceed the temperature range of –55°C to 125°C, the humidity range of 0% RH to 100% RH, or the illumination range of 0.01 Lux to 100,000 Lux are marked as range anomalies and missing, and the anomaly log is recorded. The module then calculates the difference between two consecutive valid sampling values. When the temperature change exceeds 5°C, the humidity change exceeds 15% RH, or the illumination change exceeds 2000 Lux, the point is considered a transient jump and is filled in using linear interpolation between the two previous and next points. The jump event is also recorded. If three consecutive samples are abnormal or missing, the value is backfilled based on the most recent valid value and the historical window average at a weighted ratio of 70:30. If the limit is still exceeded after backfilling, a sensor fault flag is triggered. In addition, the module supports remote transmission and adaptive learning-based threshold updates, regularly collects statistics on various abnormal events, and provides trend analysis through a reporting mechanism to ensure that subsequent data processing is based on high-quality input.

[0063] The present invention integrates a multi-channel signal acquisition and edge pre-processing unit for temperature, humidity, light, and high-resolution RGB images to achieve multi-dimensional real-time monitoring and anomaly verification of the carriage environment and the surface status of agricultural products. It can eliminate noise and anomalies at the source of the data, ensuring subsequent color difference calculation and quality assessment, and significantly improving the accuracy and stability of monitoring data.

[0064] Noise suppression includes:

[0065] S211 , performing a double exponential smoothing algorithm on the temperature T, recursively updating the level and trend components according to the preset smoothing coefficient α_T = 0.2 and trend coefficient β_T = 0.1;

[0066] S212, performing one-dimensional Kalman filtering and adaptive low-pass filtering on the humidity RH in sequence, the former updating the optimal estimate in real time based on the state space model, and the latter dynamically adjusting the cutoff frequency according to the variance of the observation sequence;

[0067] S213 , performing N=7 median filtering to eliminate extreme values ​​and three-point simple averaging processing on the illumination Lux in sequence.

[0068] The data processing module performs linear compensation calibration on temperature T and humidity RH, including:

[0069] S301. Based on the difference ΔT between the initial temperature value and the reference temperature of 25°C, the humidity is corrected by the coupling coefficient α to obtain the corrected humidity RH2, which is RH2 = RH1 - α·ΔT. RH2 is limited to be maintained in the range of 0 to 100%.

[0070] S302. A secondary correction is performed on the initial temperature value using the difference ΔRH between the corrected humidity RH2 and the reference humidity 50% RH, combined with the humidity-temperature coupling coefficient β, to obtain a final temperature compensation value T2. The calculation method is T2 = T1 – β·ΔRH, and T2 is limited to a safe range of -55°C to 125°C.

[0071] S303 , adopting an iterative convergence strategy, taking T1 and RH1 as initial iteration values, and alternately performing the above temperature compensation and humidity compensation steps three times, and updating the temperature and humidity calibration results in each iteration.

[0072] In the data processing module, the linear coupled compensation calibration of temperature and humidity first corrects the humidity by a coefficient α based on the difference between the initial temperature value and a reference temperature of 25°C, limiting the result to a range of 0 to 100%. The temperature is then corrected by a coefficient β based on the difference between the corrected humidity and a reference humidity of 50% RH, limiting the result to a range of –55°C to 125°C. These two compensations are performed alternately three times, or until the temperature change is less than 0.05°C and the humidity change is less than 0.2% RH, after which convergence is achieved. The final calibration value is then output. Furthermore, the module monitors the statistical characteristics of ΔT and ΔRH during operation and re-estimates the coupling coefficients α and β during off-peak hours using the least squares method. This allows for adaptive calibration coefficient updates, further improving the accuracy of environmental parameter measurements and system robustness.

[0073] Data processing performs color conversion on RGB images and selects representative areas including:

[0074] S401, converting the RGB image from the RGB color space to the HSV color space, and performing pixel-level filtering based on a preset H, S, and V threshold range to generate a preliminary segmentation mask Mask1;

[0075] S402, applying morphological opening and closing operations to Mask1 to remove isolated noise points and fill small holes;

[0076] S403, performing connected domain analysis on Mask2, sorting the connected domain pixels from large to small according to their area, and selecting the connected domain region with the largest area as the representative region Patch_i;

[0077] S404 , convert the representative region Patch_i and the corresponding representative region Patch_i_ref ten minutes ago into Lab space respectively, and obtain (Li, ai, bi) and (Li_ref, ai_ref, bi_ref);

[0078] S405. Calculate ΔE_pixel on all pixels in the representative area Patch_i. The calculation formula is as follows:

[0079]

[0080] S406 , averaging all ΔE_pixels in the representative region Patch_i to obtain the color difference ΔE of the representative region Patch_i.

[0081] In the data processing module, the RGB image is first converted to the HSV color space, and a preliminary mask is generated based on the preset H, S, and V thresholds. Opening and closing operations are then applied to the mask in sequence to remove isolated noise points and fill small holes to obtain a refined mask. The refined mask is subjected to connected domain analysis, and several most representative areas Patch_i are selected according to their area size. The corresponding reference areas ten minutes ago are respectively intercepted. The system converts the RGB sub-image of each area through the standard sRGB→CIELab method, calculates the color difference ΔE_pixel on all pixels, and averages the ΔE_pixel in each area to obtain the regional color difference ΔE_i. Finally, multiple ΔE_i can be weighted and synthesized into the overall color difference ΔE according to the regional area weight, providing accurate color change characteristics for subsequent quality evaluation.

[0082] The quality scoring module performs range mapping and normalization on temperature T, humidity RH, light Lux, and color difference ΔE to between 0 and 1; and performs weighted summation on each feature according to the preset weight vector to obtain the original quality score S_raw.

[0083] In the quality scoring module, temperature T, humidity RH, light Lux, and color difference ΔE are first linearly mapped to their respective preset minimum and maximum values ​​and truncated to the [0,1] interval to generate normalized features T', RH', Lux', and ΔE'. At the same time, out-of-bounds values ​​are logged.

[0084] The quality scoring module constructs an ARIMA–LSTM hybrid model including:

[0085] S501, arranging temperature T, humidity RH, light Lux, color difference ΔE, and original mass score S_raw in time series to obtain an original feature sequence;

[0086] S502, performing difference preprocessing on the original feature sequence to obtain the Sraw sequence to remove non-stationary components, and the difference order is determined by the unit root test ADF;

[0087] S503, using the autocorrelation function ACF and the partial autocorrelation function PACF of the Sraw sequence to determine the ARIMA order (p, d, q);

[0088] S504, applying the maximum likelihood estimation method to fit the ARIMA model to obtain a linear prediction function;

[0089] S505: Repeat the above steps to sequentially predict the quality index sequence within the next 6 hours.

[0090] In the quality scoring module, the five-dimensional features of temperature, humidity, light, color difference and original quality score are first constructed into a multivariate matrix according to the time series, and the unit root test and difference are performed on the quality score sequence to remove non-stationary components; then the (p, d, q) parameters are determined based on the ACF / PACF of the differenced sequence, and the ARIMA model is fitted using maximum likelihood estimation to obtain linear prediction and residual series; then the multi-source features and residuals are jointly input into the multi-layer LSTM network in a sliding window manner, and online training and fine-tuning are performed to predict the residual at the next moment; finally, the quality index QI at the next moment is generated by adding the ARIMA linear prediction and the LSTM residual prediction and inversely differencing them, and on this basis, a rolling iteration is performed to output the quality prediction series for the next 6 hours, thereby achieving high-precision, multi-step prediction of the quality dynamics during the transportation of agricultural products.

[0091] The route selection module receives a six-hour quality prediction sequence and obtains real-time traffic information for each candidate delivery route from the navigation map tool, including estimated travel time, road congestion factor, and road section safety factor. Based on the arrival time of each route, the module extracts the quality index corresponding to that arrival time from the prediction sequence. The module compares the arrival quality index of each route with a preset warning threshold, marks routes below the threshold as risky routes, and sends the risky routes to the driver.

[0092] The route selection module first receives a quality prediction sequence for the next six hours and obtains the travel time, congestion level and safety factor of each candidate route from the map service in real time. The module maps the estimated arrival time of each route to the prediction sequence, extracts the corresponding quality index, and compares it with the warning threshold. Routes with quality below the threshold are marked as risky routes and sorted by risk level to notify the driver. For routes not marked as risky, the module will also comprehensively consider the scores of the three dimensions of arrival quality, safety and timeliness, calculate the comprehensive utility through pre-set weights, and select the preferred route with the highest score; if all routes are risky, the one with the lowest risk will be automatically selected. During the entire delivery process, the module continuously monitors real-time quality and road condition changes. When the quality prediction or road condition changes suddenly cause the preferred route to no longer be optimal, it will immediately re-evaluate and push the latest driving suggestions to the driver.

[0093] Workflow: When the system is running, the multi-channel signal acquisition unit first acquires the temperature, humidity, lighting and RGB images in parallel in the vehicle cabin, and performs preprocessing such as cross-border detection, mutation interpolation and multi-stage filtering in the edge computing unit; then, the data processing module performs coupled linear compensation calibration on the temperature and humidity signals, and converts the image to HSV / CIELab space to extract representative areas and calculate color difference; the quality scoring module then normalizes and weights the calibrated environmental characteristics and color difference to generate an instantaneous quality score, and constructs an ARIMA-LSTM hybrid model to roll-forward the quality index sequence for the next 6 hours; finally, the route selection module extracts the corresponding index from the predicted sequence based on the estimated arrival time of each route, compares it with the warning threshold, marks the risky route, and alerts the driver in real time, so that the itinerary can be dynamically adjusted or preservation measures can be taken.

[0094] Beneficial effects: The present invention constructs an ARIMA-LSTM hybrid model in the quality scoring module, and uses historical temperature, humidity, light, color difference and real-time quality score multi-source information for processing and modeling analysis to achieve high-precision prediction of agricultural product quality in the next 6 hours. Based on the dynamic comparison of the prediction results with the preset threshold, it provides a decision-making basis for route selection, effectively improving the foresight and reliability of quality prediction.

[0095] The present invention monitors the quality index and warning threshold of the arrival time in the predicted quality sequence in real time, and reminds the driver in combination with the delivery time. It can issue a warning in time before the quality of agricultural products falls below the safety line, guide the vehicle to adjust the itinerary or take preservation measures, thereby minimizing quality losses and logistics costs during transportation, and improving overall delivery efficiency and product safety.

[0096] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0097] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A fresh agricultural product transportation quality monitoring system for in-transit logistics, characterized by: include: Multi-channel signal acquisition unit, edge computing unit, data processing module, quality scoring module, route selection module; The multi-channel signal acquisition module integrates sensors including a temperature probe, a capacitive humidity sensor, a photoresistor, and a high-resolution RGB camera inside the vehicle. The sensors are responsible for collecting the internal temperature T, humidity RH, and light Lux, and transmitting the RGB image to the edge computing module. The edge computing module pre-processes temperature T, humidity RH, light Lux and RGB images; The data processing module performs linear compensation calibration for temperature T and humidity RH; performs color conversion on the RGB image and selects a representative area, and calculates the color difference ΔE of the agricultural product surface ten minutes before and after in the representative area; The quality scoring module is responsible for weighted comprehensive evaluation and outputs the quality index QI. It uses temperature T, humidity RH, light Lux, color difference ΔE, and the quality index QI to build an ARIMA-LSTM hybrid model. The ARIMA-LSTM hybrid model predicts the quality prediction sequence within the next 6 hours. The route selection module monitors the quality index QI of the arrival time in the quality prediction sequence. When the quality index QI of the monitored arrival time falls below the preset threshold Q_warn, the delivery driver is reminded based on the delivery time.

2. The in-transit logistics fresh agricultural product transportation quality monitoring system according to claim 1 is characterized in that: The edge computing module performs raw data anomaly detection and noise suppression; The raw data anomaly detection includes: S201. Compare each sampling value with the upper and lower limits of the range. If the temperature T is not within the range of -55°C to 125°C, the humidity RH is not within the range of 0%RH to 100%RH, or the illumination Lux is not within the range of 0.01Lux to 100,000Lux, mark the sampling value as abnormal and set it to missing. S202, calculate the difference ΔT, ΔRH, and ΔLux between two adjacent samplings. When |ΔT|>5°C, |ΔRH|>15%RH, or |ΔLux|>2000Lux, treat the corresponding sampling value as a transient jump and fill it with the previous and next valid sampling values ​​in a linear interpolation manner, and record the jump event at the same time; S203. When three consecutive samples are abnormal or missing, backfill according to the ratio of 70%:30% based on the most recent valid value and the average value of the historical window.

3. The in-transit logistics fresh agricultural product transportation quality monitoring system according to claim 1 is characterized in that: The noise suppression includes: S211 , performing a double exponential smoothing algorithm on the temperature T, recursively updating the level and trend components according to the preset smoothing coefficient α_T = 0.2 and trend coefficient β_T = 0.1; S212, performing one-dimensional Kalman filtering and adaptive low-pass filtering on the humidity RH in sequence, the former updating the optimal estimate in real time based on the state space model, and the latter dynamically adjusting the cutoff frequency according to the variance of the observation sequence; S213 , performing N=7 median filtering to eliminate extreme values ​​and three-point simple averaging processing on the illumination Lux in sequence.

4. The in-transit logistics fresh agricultural product transportation quality monitoring system according to claim 1 is characterized in that: The data processing module performs linear compensation calibration on temperature T and humidity RH, including: S301, based on the difference ΔT between the initial temperature value and the reference temperature of 25°C, the humidity is corrected by the coupling coefficient α to obtain the corrected humidity RH2, the formula is RH2 = RH1 – α·ΔT, and RH2 is limited to be maintained in the range of 0 to 100%; S302: Using the difference ΔRH between the corrected humidity RH2 and the reference humidity of 50% RH, combined with the humidity-temperature coupling coefficient β, a secondary correction is performed on the initial temperature value to obtain a final temperature compensation value T2. The calculation method is T2 = T1 – β·ΔRH, and T2 is limited to a safe range of -55°C to 125°C. S303 , adopting an iterative convergence strategy, taking T1 and RH1 as initial iteration values, and alternately performing the above temperature compensation and humidity compensation steps three times, and updating the temperature and humidity calibration results in each iteration.

5. The in-transit logistics fresh agricultural product transportation quality monitoring system according to claim 1 is characterized in that: The data processing performs color conversion on the RGB image and selects a representative area, including: S401, converting the RGB image from the RGB color space to the HSV color space, and performing pixel-level filtering based on a preset H, S, and V threshold range to generate a preliminary segmentation mask Mask1; S402, applying morphological opening and closing operations to Mask1 to remove isolated noise points and fill small holes; S403, performing connected domain analysis on Mask2, sorting the connected domain pixels from large to small according to their area, and selecting the connected domain region with the largest area as the representative region Patch_i; S404 , convert the representative region Patch_i and the corresponding representative region Patch_i_ref ten minutes ago into Lab space respectively, and obtain (Li, ai, bi) and (Li_ref, ai_ref, bi_ref); S405. Calculate ΔE_pixel on all pixels in the representative area Patch_i. The calculation formula is as follows: S406 , averaging all ΔE_pixels in the representative region Patch_i to obtain the color difference ΔE of the representative region Patch_i.

6. The in-transit logistics fresh agricultural product transportation quality monitoring system according to claim 1 is characterized in that: The quality scoring module performs range mapping and normalization on temperature T, humidity RH, light Lux, and color difference ΔE to between 0 and 1; and performs weighted summation on each feature according to a preset weight vector to obtain the original quality score S_raw.

7. The in-transit logistics fresh agricultural product transportation quality monitoring system according to claim 1 is characterized in that: The quality scoring module constructs an ARIMA-LSTM hybrid model including: S501, arranging temperature T, humidity RH, light Lux, color difference ΔE, and original mass score S_raw in time series to obtain an original feature sequence; S502, performing difference preprocessing on the original feature sequence to obtain the Sraw sequence to remove non-stationary components, and the difference order is determined by the unit root test ADF; S503, using the autocorrelation function ACF and the partial autocorrelation function PACF of the Sraw sequence to determine the ARIMA order (p, d, q); S504, applying the maximum likelihood estimation method to fit the ARIMA model to obtain a linear prediction function; S505: Repeat the above steps to sequentially predict the quality index sequence within the next 6 hours.

8. The in-transit logistics fresh agricultural product transportation quality monitoring system according to claim 1 is characterized in that: The route selection module receives a six-hour quality prediction sequence and obtains real-time traffic information for each candidate delivery route from a navigation map tool, including estimated travel time, road congestion factor, and road section safety factor. Based on the arrival time of each route, the module extracts the quality index corresponding to that arrival time from the prediction sequence. The module compares each route's arrival quality index with a preset warning threshold, marks routes below the threshold as risky routes, and sends these risky routes to the driver.

Images