A dry bulk cargo storage abnormal state intelligent identification and low false alarm grading early warning method

Abstract

The application discloses a kind of dry bulk cargo storage abnormal state intelligent identification and low false alarm grading early warning method, it is related to dry bulk cargo storage safety monitoring and intelligent early warning technical field, including obtaining the goods type and physicochemical property of target dry bulk cargo, obtain the climatic condition in storage space, the goods type and physicochemical property and climatic condition are handled by random forest algorithm, obtain interaction feature, according to interaction feature, obtain the environmental fluctuation data in storage space, carry out multi-scale decomposition to environmental fluctuation data, and combined with goods hygroscopicity and heat release constraint input attention enhancement type long short-term memory network, obtain goods state sequence;The dry bulk cargo storage abnormal state intelligent identification and low false alarm grading early warning method provide more accurate, timely and hierarchical risk disposal basis for dry bulk cargo storage safety management.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring and intelligent early warning technology for dry bulk cargo storage, specifically to a method for intelligent identification and low false alarm graded early warning of abnormal states in dry bulk cargo storage. Background Technology

[0002] Dry bulk cargo storage is a crucial link in the logistics and turnover of bulk commodities such as grains, ores, coal, and fertilizers. Due to the characteristics of dry bulk cargo—typically large quantities, long storage periods, and difficulty in directly observing its internal condition—its storage safety is easily affected by both the cargo's own physicochemical properties and the climatic conditions of the storage space. For example, some agricultural dry bulk cargoes are highly hygroscopic; when humidity increases, ventilation conditions change, or local temperature fluctuates in the storage space, risks such as moisture absorption, heat generation, and mold growth may occur. Some coal-type dry bulk cargoes may undergo exothermic oxidation under specific temperature and humidity conditions, potentially creating a risk of spontaneous combustion.

[0003] Current safety management methods for dry bulk cargo storage typically rely on manual inspections, single-point sensor monitoring, and fixed threshold alarms. These methods often pre-set safety boundaries such as temperature, humidity, or gas concentration based on historical experience, triggering an alarm when monitored data exceeds these fixed thresholds. While relatively simple to implement, this approach primarily relies on a single environmental indicator or static alarm standard, making it difficult to accurately reflect the interaction between different cargo types, their physical and chemical properties, and storage climate conditions. In actual storage, abnormal states in dry bulk cargo are often not directly caused by a single abrupt change in an indicator, but rather by the combined effects of cargo properties, environmental fluctuations, and the cumulative effect over time. For cargoes with significant hygroscopic and exothermic characteristics, even if the temperature or humidity in the storage space has not yet exceeded traditional alarm thresholds, a risk evolution trend may already be emerging within the cargo. If fixed safety boundaries are still used for judgment, it can easily lead to delayed identification of abnormal states, or even missed warning signals. Furthermore, environmental fluctuations within the storage space are unstable and multi-scale, with short-term disturbances, periodic changes, and continuous risk evolution potentially coexisting. Existing alarm methods often struggle to distinguish between short-term, harmless fluctuations and the actual evolution of risks, easily misjudging occasional environmental disturbances as dangerous conditions, resulting in a high false alarm rate. Frequent false alarms not only increase manual handling costs but also weaken the reliability of the early warning system, impacting the efficiency of warehouse safety management. Furthermore, existing early warning standards typically lack the ability to dynamically update with real-time weather conditions and cargo status changes, making it difficult to classify abnormal states according to the degree of risk evolution. When dry bulk cargo gradually transitions from a normal state to a high-risk state, traditional methods often only provide a single alarm result, failing to accurately distinguish different risk levels and providing warehouse managers with a tiered, low-false-alarm-rate response strategy. Summary of the Invention

[0004] The purpose of this invention is to provide a method for intelligent identification and low false alarm tiered early warning of abnormal states in dry bulk cargo storage, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent identification and low false alarm graded early warning of abnormal states in dry bulk cargo storage, comprising: S1, acquiring the cargo type and physicochemical properties of the target dry bulk cargo, acquiring the climatic conditions within the storage space, processing the cargo type, physicochemical properties, and climatic conditions to obtain interaction characteristics, including interaction characteristics between moisture absorption rate and relative humidity change, interaction characteristics between moisture content and temperature change rate, interaction characteristics between porosity and ventilation velocity, and interaction characteristics between moisture absorption and heat release coefficient and dew point difference; S2, based on the interaction characteristics, obtaining... S3. Obtain environmental fluctuation data within the storage space, perform multi-scale decomposition on the environmental fluctuation data to obtain a cargo state sequence; S4. Acquire historical experience data, and determine the initial safety boundary by comparing the cargo state sequence with the historical experience data; S5. If there are cargo state sequences within the initial safety boundary that deviate from the historical experience data, extract the risk evolution characteristics corresponding to the cargo state sequences that deviate from the historical experience data; S6. Obtain the initial warning standard, and obtain the first critical value by calculating the deviation value between the risk evolution characteristics and the initial warning standard. If the first critical value exceeds a preset threshold, then the first critical number is determined. The cargo status sequence corresponding to the value is the first danger signal. The initial warning standard is obtained by comparing the risk evolution characteristics with the historical safety baseline. S6. For the first danger signal, real-time climate conditions are obtained. By fusing the first danger signal and the real-time climate conditions, a second critical value is obtained. If the second critical value is greater than the first critical value, the second danger signal corresponding to the second critical value is extracted. S7. The second danger signal is confirmed. If the second danger signal continuously meets the preset confirmation conditions within a preset time window, the second danger signal is retained. If the second danger signal does not continuously meet the preset confirmation conditions, the second danger signal is filtered or downgraded. S8. According to the retained second danger signal, the initial warning standard is obtained. By replacing the first critical value in the initial warning standard with the second critical value, the updated warning standard is obtained. The initial warning standard formed and saved in S5 is read. The initial warning standard includes an observation interval, a warning interval, and a danger interval. The average value of the second critical value in the retained second danger signal is compared with the above three intervals one by one. The interval where the average value of the second critical value is located is determined as the interval to be updated. After the interval to be updated is determined, the original first critical value in the interval is read.

[0006] S1 includes:

[0007] Acquire the cargo type and initial physicochemical properties data of the target dry bulk cargo, clean the initial physicochemical properties data to obtain standard physicochemical properties characteristics, and acquire the initial climate conditions data within the storage space;

[0008] A feature extraction algorithm is used to process standard physicochemical properties and initial climate conditions data to obtain a multidimensional environmental cargo feature matrix.

[0009] If the feature dimension of the multidimensional environmental cargo feature matrix exceeds the preset dimension threshold, the multidimensional environmental cargo feature matrix is ​​reduced in dimension by principal component analysis algorithm to obtain the core feature vector.

[0010] The input dataset is constructed based on the core feature vectors. The random forest algorithm is used to perform node splitting calculations on the input dataset to obtain the interaction characteristics of cargo types, physicochemical properties, and climatic conditions of the target dry bulk cargo.

[0011] S2 includes:

[0012] Obtain an environmental fluctuation data matrix based on interaction characteristics;

[0013] The environmental fluctuation data matrix is ​​decomposed to obtain multi-scale environmental decomposition components.

[0014] The multi-scale environment is decomposed and its components are reduced in dimensionality to obtain the core environmental feature vector, and the hygroscopic and exothermic physical properties are obtained to construct boundary constraint conditions.

[0015] An input tensor is constructed based on the core environmental feature vector and boundary constraints. The input tensor is then processed through a long short-term memory network to obtain the cargo state sequence.

[0016] Preferably, step S1 includes acquiring the cargo type and initial physicochemical properties data of the target dry bulk cargo, cleaning the initial physicochemical properties data to obtain standard physicochemical property characteristics, and acquiring the initial climate condition data within the storage space; processing the standard physicochemical property characteristics and initial climate condition data using a feature extraction algorithm to obtain a multidimensional environmental cargo feature matrix; if the feature dimension of the multidimensional environmental cargo feature matrix exceeds a preset dimension threshold, then performing dimensionality reduction processing on the multidimensional environmental cargo feature matrix using a principal component analysis algorithm to obtain a core feature vector; constructing an input dataset based on the core feature vector, and performing node splitting calculation on the input dataset using a random forest algorithm to obtain the interaction characteristics of the cargo type, physicochemical properties, and climate conditions of the target dry bulk cargo.

[0017] Preferably, step S2 includes obtaining an environmental fluctuation data matrix based on interaction characteristics; decomposing the environmental fluctuation data matrix to obtain multi-scale environmental decomposition components; reducing the dimensionality of the multi-scale environmental decomposition components to obtain a core environmental feature vector, and obtaining hygroscopic and exothermic physical properties to construct boundary constraints; constructing an input tensor based on the core environmental feature vector and boundary constraints, and processing the input tensor through a long short-term memory network to obtain a cargo state sequence.

[0018] Preferably, step S3 includes acquiring a cargo state sequence, calculating the trajectory similarity between the cargo state sequence and historical experience data; determining target historical experience data based on the trajectory similarity, comparing the cargo state sequence with the target historical experience data to obtain the state fluctuation amplitude; if the experience matching degree corresponding to the state fluctuation amplitude is lower than a threshold, generating an abnormal state node; calculating the boundary deviation value based on the abnormal state node, and determining the initial safety boundary based on the boundary deviation value.

[0019] Preferably, step S4 includes calculating the difference between the cargo state sequence deviating from historical experience data and the baseline state to obtain a state offset vector; obtaining the temporal evolution gradient based on the state offset vector, fusing the temporal evolution gradient with environmental factors to determine the risk evolution rate; if the risk evolution rate is greater than a threshold, constructing a trend matrix based on the risk evolution rate; and performing dimensionality reduction processing on the trend matrix to obtain the risk evolution characteristics corresponding to the cargo state sequence deviating from historical experience data.

[0020] Preferably, step S5 includes obtaining an initial warning standard, which is obtained by comparing risk evolution characteristics with a historical safety baseline; calculating the projection distance of the risk evolution characteristics within the corresponding interval of the initial warning standard to obtain a deviation value; obtaining a first critical value based on the deviation value; determining whether the first critical value exceeds a preset threshold; and if the first critical value exceeds the preset threshold, determining the cargo state sequence corresponding to the first critical value as a first danger signal.

[0021] Preferably, step S6 includes obtaining the initial risk level corresponding to the first danger signal, obtaining real-time climate conditions based on the initial risk level; performing dimensionality reduction processing on the real-time climate conditions to obtain core meteorological data, determining fusion weights based on the core meteorological data; fusing the first danger signal and the core meteorological data through the fusion weights to obtain a second critical value; determining whether the second critical value is greater than the first critical value; and if the second critical value is greater than the first critical value, extracting the second danger signal corresponding to the second critical value.

[0022] Preferably, step S7 includes acquiring a second danger signal and extracting a feature sequence, processing the feature sequence within a preset time window to obtain signal strength and fluctuation frequency; calculating the attenuation amplitude based on the signal strength and fluctuation frequency, determining the duration of the second danger signal based on the attenuation amplitude; determining whether the duration meets a preset confirmation condition; if the duration meets the preset confirmation condition, retaining the second danger signal; if the duration does not meet the preset confirmation condition, filtering or downgrading the second danger signal.

[0023] Preferably, step S8 includes acquiring the retained second danger signal, extracting features of the second danger signal to obtain an initial warning standard; calculating the deviation between the first critical value and the features within the initial warning standard to determine a second critical value; replacing the first critical value with the second critical value, acquiring the initial warning standard based on the retained second danger signal, and obtaining an updated warning standard by replacing the first critical value in the initial warning standard with the second critical value.

[0024] Preferably, the method further includes S9: determining the storage anomaly level of the target dry bulk cargo based on the degree of deviation between the second critical value and the updated warning standard, and outputting graded warning information corresponding to the storage anomaly level. Specifically, this includes acquiring storage environment data; determining the second critical value of the target dry bulk cargo based on the storage environment data; calculating the degree of deviation between the second critical value and the updated warning standard, wherein the updated warning standard is obtained by fitting historical warning records; determining whether the degree of deviation is greater than a preset deviation threshold; if the degree of deviation is greater than the preset deviation threshold, determining the storage anomaly level of the target dry bulk cargo based on the degree of deviation, and outputting graded warning information corresponding to the storage anomaly level.

[0025] As can be seen from the above technical solution, the present invention has the following beneficial effects:

[0026] This intelligent identification and low-false-alarm tiered early warning method for abnormal states in dry bulk cargo storage acquires information on the type, physicochemical properties, and climatic conditions within the storage space of the target dry bulk cargo. It then utilizes a random forest algorithm to extract the interaction features between cargo attributes and climatic conditions. This allows early warning judgments to no longer rely on single environmental indicators or fixed empirical thresholds, but rather comprehensively reflect the varying sensitivities of different dry bulk cargoes to changes in the storage environment. By performing multi-scale decomposition of environmental fluctuation data and combining it with a cargo moisture absorption and heat release constraint-input attention-enhanced long short-term memory network, it can more accurately identify the evolution trend of cargo states over time, improving the early detection capability of potential abnormal states such as internal heating, moisture absorption and mold growth, and oxidative heat release. Finally, by comparing the cargo state sequence with historical experience data to determine the initial safety boundary, and addressing deviations as needed... By extracting risk evolution characteristics from historical experience data state sequences, abnormal signals that have not yet exceeded traditional alarm thresholds but already show a risk development trend can be identified, reducing the possibility of missed alarm signals. By fusing the first danger signal with real-time climate conditions to obtain a second critical value, and continuously confirming the second danger signal, non-persistent abnormal signals caused by short-term environmental disturbances can be filtered or downgraded, thereby reducing the false alarm rate. By replacing the first critical value in the initial warning standard with the second critical value, the warning standard can be dynamically updated according to the cargo status and real-time climate conditions. The level of abnormal storage status can be determined based on the degree of deviation between the second critical value and the updated warning standard, and corresponding graded warning information can be output, thus providing a more accurate, timely, and hierarchical basis for risk handling in dry bulk storage safety management. Attached Figure Description

[0027] Figure 1 This is a signal transmission diagram of the present invention;

[0028] Figure 2 This is a structural block diagram of the local terminal of an exemplary electronic device of the present invention;

[0029] Figure 3 This is a structural block diagram of the network terminal of an exemplary electronic device of the present invention. Detailed Implementation

[0030] 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.

[0031] Example 1:

[0032] like Figure 1As shown, this invention provides a technical solution: a method for intelligent identification and low false alarm tiered early warning of abnormal states in dry bulk cargo storage, comprising:

[0033] S1. Obtain the cargo type and physicochemical properties of the target dry bulk cargo, obtain the climate conditions in the storage space, and process the cargo type, physicochemical properties and climate conditions through the random forest algorithm to obtain the interaction characteristics.

[0034] S2. Based on the interaction characteristics, obtain environmental fluctuation data within the storage space, perform multi-scale decomposition on the environmental fluctuation data, and combine it with the attention-enhanced long short-term memory network with the constraint of moisture absorption and heat release of goods to obtain the state sequence of goods.

[0035] S3. For the cargo status sequence, obtain historical experience data, and determine the initial safety boundary by comparing the cargo status sequence with the historical experience data;

[0036] S4. If there is a cargo state sequence that deviates from historical experience data within the initial safety boundary, then extract the risk evolution features corresponding to the cargo state sequence that deviates from historical experience data.

[0037] S5. Based on the risk evolution characteristics, obtain the initial warning standard. By calculating the deviation between the risk evolution characteristics and the initial warning standard, obtain the first critical value. If the first critical value exceeds the preset threshold, then determine the cargo state sequence corresponding to the first critical value as the first danger signal.

[0038] S6. For the first danger signal, obtain real-time climate conditions, and obtain a second critical value by fusing the first danger signal and the real-time climate conditions. If the second critical value is greater than the first critical value, then extract the second danger signal corresponding to the second critical value.

[0039] S7. Continuously confirm the second danger signal. If the second danger signal continuously meets the preset confirmation conditions within the preset time window, the second danger signal is retained. If the second danger signal does not continuously meet the preset confirmation conditions, the second danger signal is filtered or downgraded.

[0040] S8. Based on the retained second danger signal, obtain the initial warning standard, and obtain the updated warning standard by replacing the first critical value in the initial warning standard with the second critical value.

[0041] S9. Based on the degree of deviation between the second critical value and the updated warning standard, determine the storage anomaly level of the target dry bulk cargo, and output the graded warning information corresponding to the storage anomaly level.

[0042] In this embodiment, the target dry bulk cargo is one of the following: coal, iron ore, grain, fertilizer, sand and gravel, and cement clinker. The system obtains the cargo type of the target dry bulk cargo through the warehouse management system and obtains the physical and chemical properties of the target dry bulk cargo through the inbound inspection equipment.

[0043] Specifically, the physicochemical properties include moisture content, particle size distribution, bulk density, porosity, moisture absorption rate, thermal conductivity, specific heat capacity, moisture absorption and heat release coefficient, critical moisture content, and angle of repose. The system acquires the climatic conditions in the storage space through temperature and humidity sensors, dew point sensors, wind speed sensors, air pressure sensors, and ventilation equipment status acquisition modules deployed in the storage space. The climatic conditions include the temperature inside the storage space, the relative humidity inside the storage space, the dew point temperature, the ventilation wind speed, the temperature difference between inside and outside the storage space, and the air pressure value.

[0044] Furthermore, the random forest algorithm is used to perform nonlinear correlation analysis on cargo types, physicochemical properties, and climatic conditions.

[0045] Specifically, the system first classifies the types of goods by coding them, then normalizes the moisture content, bulk density, porosity, moisture absorption rate, thermal conductivity, specific heat capacity, warehouse temperature, warehouse relative humidity, dew point temperature, ventilation velocity, temperature difference between inside and outside the warehouse, and air pressure. The processed data is then input into a random forest model. The random forest model consists of 100 decision trees, each of which performs splitting judgments on the input features based on the Gini index or mean square error criterion. The random forest model outputs the interaction features between the goods' attributes and climatic conditions. These interaction features include the interaction between moisture absorption rate and relative humidity change, moisture content and temperature change rate, porosity and ventilation velocity, and the interaction between the moisture absorption and heat release coefficient and the dew point difference.

[0046] The aforementioned interaction characteristics are used to characterize the risk basis of target dry bulk cargoes experiencing dampness, condensation, caking, abnormal temperature changes, or heat accumulation under changes in the storage environment.

[0047] Furthermore, the system determines the objects to be collected for environmental fluctuation data based on the characteristics of the interaction.

[0048] Specifically, the environmental fluctuation data includes temperature, humidity, dew point difference, ventilation speed, temperature change rate, and humidity change rate sequences at different measuring points within the storage space. Each measuring point forms a time series with a sampling period of 5 minutes. The system performs a 3-scale decomposition on the environmental fluctuation data, which corresponds to a multi-scale decomposition. The 3-scale decomposition includes short-period, medium-period, and long-period components. The short-period component is used to characterize instantaneous disturbances within 0 to 30 minutes, the medium-period component is used to characterize trend changes within 30 minutes to 6 hours, and the long-period component is used to characterize cumulative changes within 6 hours to 72 hours. The system transforms the original environmental fluctuation data into environmental change characteristics at different time scales through 3-scale decomposition, enabling subsequent models to identify instantaneous fluctuations, trend shifts, and cumulative risks respectively.

[0049] In this embodiment, the moisture absorption and heat release constraint of the cargo is used to introduce the moisture-heat coupling change mechanism of the target dry bulk cargo into the state recognition process. When the relative humidity in the warehouse reaches 85% and the difference between the dew point temperature and the surface temperature of the cargo is less than 2°C, the system determines that the target dry bulk cargo has moisture absorption conditions.

[0050] Specifically, when the moisture content of the target dry bulk cargo increases by 0.3% within 1 hour and the ventilation velocity is less than 0.5 m / s, the system determines that the cargo's moisture absorption process has the conditions for heat accumulation. The system uses the change in moisture content, moisture absorption rate, moisture absorption and heat release coefficient, thermal conductivity, specific heat capacity, and ventilation velocity as inputs to the cargo's moisture absorption and heat release constraints, and inputs them together with the environmental fluctuation data after 3-scale decomposition into the attention-enhanced long short-term memory network.

[0051] Specifically, the attention-enhanced long short-term memory network includes an input layer, a long short-term memory layer, an attention weight layer, and an output layer.

[0052] The system comprises the following layers: an input layer receives 3-scale environmental fluctuation features and cargo moisture absorption and heat release constraint features; a long short-term memory layer extracts continuous change relationships from the time series; an attention weight layer assigns weights to different sampling times, measurement points, and environmental variables; and an output layer generates a cargo state sequence, which includes a continuous risk score sequence and a state label sequence. The continuous risk score sequence ranges from 0 to 1; the state label sequence includes normal state, observation state, abnormal evolution state, and dangerous state. A risk score of 0 to 0.3 corresponds to a normal state; a risk score greater than 0.3 and less than or equal to 0.5 corresponds to an observation state; a risk score greater than 0.5 and less than or equal to 0.75 corresponds to an abnormal evolution state; and a risk score greater than 0.75 and less than or equal to 1 corresponds to a dangerous state.

[0053] In this embodiment, historical experience data includes normal storage records, abnormal storage records, disposal records, inspection records, and existing early warning records formed under the same storage type for the same type of goods. The system compares the currently obtained goods status sequence with the historical experience data to determine the initial safety boundary.

[0054] The initial safety boundary is determined by the average and standard deviation of the state scores of historical normal state samples.

[0055] Specifically, the system first calculates the average state score of historical normal state samples, then calculates the difference between the state score of each historical normal state sample and the average value, and then calculates the standard deviation based on all differences. The system uses the value obtained by subtracting twice the standard deviation from the average state score as the lower limit of the initial safety boundary, and the value obtained by adding twice the standard deviation to the average state score as the upper limit of the initial safety boundary. When the current state score is between the lower limit and the upper limit, the system includes the cargo state sequence corresponding to the state score in the initial safety boundary. When the current state score is less than the lower limit or greater than the upper limit, the system marks the cargo state sequence corresponding to the state score as an out-of-boundary abnormal sequence.

[0056] Furthermore, when the cargo state sequence is within the initial safety boundary, but its change pattern deviates from the normal pattern in historical experience data, the system extracts the risk evolution characteristics corresponding to the cargo state sequence.

[0057] Specifically, the risk evolution characteristics include deviation magnitude, deviation duration, deviation growth rate, cumulative deviation, fluctuation energy, trend slope, anomaly point expansion rate, and temperature-humidity coupling strength. Deviation magnitude is determined by the difference between the current state score and the average historical normal state score. Deviation duration is determined by multiplying the number of consecutive deviation sampling points by the sampling period. Deviation growth rate is determined by dividing the difference in deviation magnitudes between adjacent sampling points by the sampling interval. Cumulative deviation is determined by summing the deviation magnitudes of all sampling points within a preset time window. Fluctuation energy is determined by summing the squares of the deviation magnitudes of all sampling points within a preset time window. The trend slope is determined by linearly fitting the state score sequence within the preset time window. The anomaly point expansion rate is determined by the number of newly added anomaly sampling points per unit time. The temperature-humidity coupling strength is determined by the correlation between the temperature change rate sequence and the humidity change rate sequence.

[0058] As a preferred embodiment, the system obtains an initial warning standard based on the risk evolution characteristics and calculates the deviation value between the risk evolution characteristics and the initial warning standard to obtain a first critical value.

[0059] Specifically, the first critical value is determined by a weighted sum of five indicators. The system first obtains the deviation magnitude, deviation duration, deviation growth rate, temperature and humidity coupling strength, and abnormal measurement point expansion rate, respectively. Then, it normalizes each of the five indicators to make their values ​​range from 0 to 1. Next, it multiplies the normalized result corresponding to the deviation magnitude by 0.3, the normalized result corresponding to the deviation duration by 0.2, the normalized result corresponding to the deviation growth rate by 0.2, the normalized result corresponding to the temperature and humidity coupling strength by 0.2, and the normalized result corresponding to the abnormal measurement point expansion rate by 0.1. Finally, it adds the five products to obtain the first critical value.

[0060] The initial warning criteria include an observation threshold of 0.5, an anomaly threshold of 0.75, and a danger threshold of 0.9. When the first critical value is greater than 0.5, the system determines that the cargo status sequence corresponding to the first critical value is the first danger signal.

[0061] Furthermore, in response to the first danger signal, the system acquires real-time climate conditions and fuses the first danger signal with the real-time climate conditions to obtain a second critical value.

[0062] Specifically, real-time climate conditions include real-time warehouse temperature, real-time relative humidity, real-time dew point temperature, real-time ventilation speed, real-time temperature difference between inside and outside the warehouse, and real-time air pressure. The system determines a risk correction coefficient based on the real-time climate conditions and multiplies the first critical value by the risk correction coefficient to obtain the second critical value. When the three conditions of real-time relative humidity reaching 85%, real-time ventilation speed being less than 0.5 m / s, and the difference between real-time dew point temperature and cargo surface temperature being less than 2℃ are simultaneously met, the risk correction coefficient is 1.2; when any two of the above three conditions are simultaneously met, the risk correction coefficient is 1.1; when only one of the above three conditions is met, the risk correction coefficient is 1.05; when none of the above three conditions are met, the risk correction coefficient is 1. If the second critical value is greater than the first critical value, the system extracts the second danger signal corresponding to the second critical value.

[0063] In this embodiment, the system continuously confirms the second danger signal. The continuous confirmation is performed within a preset time window of 2 hours, and the sampling period is 5 minutes. Therefore, each preset time window includes 24 sampling points.

[0064] Specifically, the preset confirmation conditions include: the second critical value is greater than 0.5 in at least 20 out of 24 sampling points; the linear fitting slope of the risk score sequence is greater than 0 within 2 hours; the number of abnormal measurement points does not decrease within 2 hours; and the average value of the temperature and humidity coupling strength is greater than 0.6 within 2 hours. When the second danger signal meets the above conditions simultaneously, the system retains the second danger signal. When the second danger signal does not meet the above conditions simultaneously, the system filters or downgrades the second danger signal.

[0065] Specifically, when the second critical value is less than or equal to 0.5, the system filters the signal; when the second critical value is greater than 0.5 and less than or equal to 0.75, the system downgrades the signal to an observation signal.

[0066] In this embodiment, the system obtains the initial warning standard based on the retained second danger signal, and replaces the first critical value in the initial warning standard with the second critical value to obtain the updated warning standard.

[0067] Specifically, the updated warning criteria include an observation threshold of 0.5, an anomaly threshold of 0.75, and a danger threshold of 0.9. The system determines the level of abnormal storage status of the target dry bulk cargo based on the degree of deviation between the second critical value and the updated warning criteria. When the second critical value is greater than 0.5 and less than or equal to 0.75, it is determined as a Level 1 warning; when the second critical value is greater than 0.75 and less than or equal to 0.9, it is determined as a Level 2 warning; and when the second critical value is greater than 0.9 and less than or equal to 1, it is determined as a Level 3 warning.

[0068] Furthermore, the system outputs tiered early warning information corresponding to the level of abnormal warehouse status.

[0069] Specifically, the tiered early warning information includes the anomaly level, anomaly measurement point number, anomaly type, second critical value, deviation degree, risk evolution trend, warning time, and response measures. The response measures for a Level 1 early warning are to initiate manual verification and add one on-site inspection; the response measures for a Level 2 early warning are to activate ventilation equipment, dehumidification equipment, or perform localized turning operations; and the response measures for a Level 3 early warning are to initiate on-site emergency verification, isolate the abnormal area, and generate an emergency response work order.

[0070] Through the above process, the system first identifies the first danger signal based on the cargo status sequence, then forms the second danger signal by combining real-time weather conditions, and continuously confirms and filters instantaneous disturbance signals, thereby realizing the identification and graded early warning of abnormal states in dry bulk cargo storage.

[0071] Example 2:

[0072] S1 includes acquiring the cargo types and initial physicochemical properties data of the target dry bulk cargo, cleaning the initial physicochemical properties data to obtain standard physicochemical property characteristics, and acquiring the initial climate condition data within the storage space; using a feature extraction algorithm to process the standard physicochemical property characteristics and initial climate condition data to obtain a multidimensional environmental cargo feature matrix; if the feature dimension of the multidimensional environmental cargo feature matrix exceeds a preset dimension threshold, then the multidimensional environmental cargo feature matrix is ​​dimensionality reduced using a principal component analysis algorithm to obtain a core feature vector; constructing an input dataset based on the core feature vector, and performing node splitting calculations on the input dataset using a random forest algorithm to obtain the interaction characteristics of the cargo types, physicochemical properties, and climate conditions of the target dry bulk cargo.

[0073] In this embodiment, the type of target dry bulk cargo is determined by the warehousing registration information, which includes cargo name, batch number, warehousing time, warehouse location number, and source unit.

[0074] Specifically, the initial physicochemical properties data are obtained from the incoming inspection equipment and re-inspection records. The data items include moisture content, particle size distribution, bulk density, porosity, moisture absorption rate, thermal conductivity, specific heat capacity, moisture absorption and heat release coefficient, critical moisture content, and angle of repose. After each batch of target dry bulk cargo is put into storage, the processing unit writes the cargo type, batch number, warehouse location number, and initial physicochemical properties data into the same basic record, so that the cargo data corresponds to the storage space location.

[0075] Furthermore, when cleaning the initial physicochemical property data, field integrity verification is performed first.

[0076] Specifically, the processing unit reads the moisture content, particle size distribution, bulk density, porosity, moisture absorption rate, thermal conductivity, specific heat capacity, moisture absorption and heat release coefficient, critical moisture content, and angle of repose item by item. When a certain field is empty, non-numerical, or lacks units, the field is marked as a field to be filled. The filling of the field to be filled is performed according to the re-inspection record of the same batch and the same warehouse. Specifically, all valid test values ​​of the field in the re-inspection record are read, all valid test values ​​are added together, and then divided by the number of valid test values. The result is used as the fill value of the field.

[0077] When there is no valid test value in the re-inspection record, the valid test value of the same type of goods on the same warehousing date is read, and the missing value is obtained by averaging the same values.

[0078] In this embodiment, outlier handling is performed according to two conditions: equipment range and re-inspection difference. The processing unit first reads the lower limit and upper limit of the range of the inbound testing equipment. When the test value is less than the lower limit or greater than the upper limit, the test value is marked as an invalid test value.

[0079] For the detection values ​​within the measurement range, the processing unit reads three retest values ​​of the same field, first adds the three retest values ​​and divides them by 3 to obtain the retest average, then calculates the square of the difference between the three retest values ​​and the retest average, adds the three squared differences and divides them by 3, and then takes the square root of the result to obtain the retest standard deviation; when the absolute difference between the detection value and the retest average exceeds three times the retest standard deviation, the detection value is deleted and replaced with the retest average.

[0080] The resulting moisture content, particle size distribution, bulk density, porosity, moisture absorption rate, thermal conductivity, specific heat capacity, moisture absorption and heat release coefficient, critical moisture content, and angle of repose constitute the standard physicochemical properties.

[0081] In this embodiment, unit unification is performed after outlier processing. Moisture content and critical moisture content are uniformly converted to percentage values. When the original record is in decimal form, the original value is multiplied by 100. Particle size is uniformly converted to millimeters. When the original record is in meters, the original value is multiplied by 1000. Bulk density is uniformly converted to kilograms per cubic meter. Thermal conductivity is uniformly converted to watts per meter Kelvin. Specific heat capacity is uniformly converted to joules per kilogram Kelvin. After unit unification, each standard physicochemical property is normalized.

[0082] During normalization, the minimum and maximum values ​​of the feature in the training samples are read first. The minimum value is subtracted from the current feature value. The difference is then divided by the difference between the maximum and minimum values ​​to obtain a standardized value ranging from 0 to 1.

[0083] Furthermore, the initial climate condition data within the storage space is obtained by temperature and humidity sensors, dew point sensors, wind speed sensors, air pressure sensors, and ventilation equipment status acquisition modules.

[0084] Specifically, the data collected includes the following: warehouse temperature, relative humidity, dew point temperature, ventilation velocity, temperature difference between inside and outside the warehouse, air pressure, and the operating status of ventilation equipment. The sampling cycle is 5 minutes. The sampling record includes the sampling time, measuring point number, and warehouse number. The processing unit matches the initial climate condition data with standard physicochemical properties based on the warehouse number. When there are 4 measuring points in the same warehouse, the average temperature, average humidity, average dew point temperature, average ventilation velocity, and average air pressure of the 4 measuring points at the same sampling time are calculated, and the obtained average values ​​are written into the climate condition record of that warehouse.

[0085] Furthermore, when processing standard physicochemical properties and initial climate conditions data, the feature extraction algorithm first generates basic features, and then generates variation features, difference features, and combination features.

[0086] Specifically, the basic characteristics are directly derived from normalized moisture content, particle size distribution, bulk density, porosity, moisture absorption rate, thermal conductivity, specific heat capacity, moisture absorption and heat release coefficient, critical moisture content, angle of repose, storage temperature, storage relative humidity, dew point temperature, ventilation velocity, temperature difference between inside and outside the storage, and air pressure. Variation characteristics are calculated based on two adjacent sampling times. The processing unit subtracts the temperature of the previous sampling time from the temperature of the later sampling time, and then divides by the time interval between the two sampling times to obtain the temperature change rate. The humidity change rate, dew point change rate, and ventilation velocity change rate are obtained using the same process.

[0087] Among them, the difference features include the difference between dew point temperature and cargo surface temperature, the difference between real-time moisture content and critical moisture content, and the difference between indoor temperature and outdoor temperature. The calculation is performed by subtracting the next value from the previous value.

[0088] Specifically, the combined features are used to reflect the interaction between the physical and chemical properties of goods and climatic conditions.

[0089] Specifically, the processing unit multiplies the normalized moisture content with the normalized relative humidity in the storage room to obtain the combined characteristics of moisture content and relative humidity; multiplies the normalized moisture absorption rate with the normalized dew point difference to obtain the combined characteristics of moisture absorption rate and dew point difference; multiplies the normalized porosity with the normalized ventilation velocity to obtain the combined characteristics of porosity and ventilation velocity; and multiplies the normalized moisture absorption heat release coefficient with the normalized temperature change rate to obtain the combined characteristics of moisture absorption heat release coefficient and temperature change rate.

[0090] The above-mentioned combined features transform the common variation relationship of two individual features in the same sampling record into a numerical term that can be input into the model.

[0091] Furthermore, after obtaining the various features, the processing unit constructs an environmental cargo feature matrix based on the sampling records.

[0092] Specifically, each row of the matrix corresponds to one sampling record, and each column corresponds to one feature item. The column items include, in order, the cargo type code, standard physicochemical properties, initial climate conditions data, variation characteristics, difference characteristics, and combination characteristics. The cargo type codes are generated using a sequential coding method: coal is coded as 1, iron ore as 2, grain as 3, fertilizer as 4, sand and gravel as 5, and cement clinker as 6.

[0093] After the processing unit completes the matrix construction, it counts the number of feature columns involved in the calculation in the matrix and uses this number as the feature dimension.

[0094] Specifically, the preset dimensionality thresholds are determined through historical sample verification. The processing unit calls upon warehouse samples that have been labeled within the last 12 months and sets the candidate dimensionality thresholds to 20, 25, 30, 35, and 40. For each candidate dimensionality threshold, the processing unit retains the corresponding number of feature columns, trains a random forest model, and calculates the recognition accuracy in the verification samples.

[0095] Specifically, the accuracy rate is calculated as follows: the number of correctly identified states in the validation samples is counted, and then divided by the total number of validation samples. The processing unit also records the average computation time for a single sample. The average computation time is calculated as follows: the time taken to complete the identification of all validation samples is divided by the total number of validation samples. When the candidate dimension threshold is 30, the accuracy rate reaches 0.92, and the average computation time per sample does not exceed 200 milliseconds. When the candidate dimension threshold is less than 30, the accuracy rate does not reach 0.92. When the candidate dimension threshold is greater than 30, the accuracy rate does not increase by more than 0.01. Based on the above validation results, the dimension threshold is determined to be 30.

[0096] Furthermore, when the feature dimension of the environmental cargo feature matrix exceeds 30, the processing unit uses principal component analysis algorithm for dimensionality reduction.

[0097] Before dimensionality reduction, the processing unit first standardizes each column of features in the matrix. The standardization process is as follows: calculate the average value of all sample values ​​in the column, subtract the average value from each sample value; then calculate the standard deviation of all sample values ​​in the column, divide the aforementioned difference by the standard deviation, and obtain the standardized feature value. After standardization, each feature column has the same numerical scale, avoiding the influence of dimensionality reduction results due to differences in the dimensions of moisture content, bulk density, temperature, and wind speed.

[0098] Specifically, when the principal component analysis algorithm is executed, the processing unit calculates the covariance between each feature column based on the standardized matrix. For any two feature columns, the two standardized feature values ​​in the same row are multiplied together, the products of all rows are summed, and finally the result is divided by the number of sample rows minus 1 to obtain the covariance between the two feature columns.

[0099] After calculating the covariance between all feature columns, a covariance matrix is ​​formed. The processing unit calculates the eigenvalues ​​and eigenvectors of the covariance matrix and arranges the eigenvectors in descending order of eigenvalues. The threshold for the information retention ratio is set at 95%. The determination process is as follows: 90%, 95%, and 98% are used as candidate ratios for model validation. The validation accuracy corresponding to 95% reaches 0.92, and the average calculation time per sample does not exceed 200 milliseconds. The validation accuracy increase corresponding to 98% does not reach 0.01, and the calculation time exceeds 200 milliseconds. The validation accuracy corresponding to 90% does not reach 0.92, so 95% is selected.

[0100] Furthermore, after obtaining the sorted feature vectors, the processing unit sequentially accumulates the corresponding feature values ​​starting from the first feature vector, and divides the sum of the accumulated feature values ​​by the sum of all feature values ​​to obtain the current information retention ratio.

[0101] When the current information retention rate reaches 95%, the selection of feature vectors is stopped. Subsequently, the processing unit multiplies the standardized environmental cargo feature matrix with the selected feature vectors to obtain the dimensionality-reduced core feature vector. This core feature vector retains the change information between the cargo's physical and chemical properties and the initial climate conditions that contributes to state identification, and reduces duplicate features entering the subsequent random forest model.

[0102] As a preferred embodiment, when constructing the input dataset based on the core feature vector, the processing unit organizes each sample into a core feature vector, cargo type code, warehouse number, sampling time, and status marker.

[0103] Specifically, the status markers are derived from historical storage records, inspection records, and anomaly handling records. The normal status corresponds to samples for which no action has been taken and the inspection records show no anomalies; the observation status corresponds to samples for which the inspection records show verification but no action has been taken; the abnormal evolution status corresponds to samples for which ventilation, dehumidification, or turning measures have been taken; and the dangerous status corresponds to samples for which isolation measures or emergency work orders have been initiated. The processing units are arranged in the order of sampling time and input into the dataset. The first 70% of the samples are used as training samples, and the last 30% of the samples are used as validation samples, so that the model validation process conforms to the actual order of the evolution of storage status over time.

[0104] Furthermore, the random forest algorithm performs node splitting calculations on the input dataset.

[0105] Specifically, the random forest model includes 100 decision trees, the number of which is determined by the number of candidate trees. The number of candidate trees is 50, 100, and 150. The processing unit trains three sets of models respectively. The verification accuracy corresponding to 100 decision trees reaches 0.92, and the average computation time per sample does not exceed 200 milliseconds. The verification accuracy corresponding to 150 decision trees does not reach 0.01, and the computation time exceeds 200 milliseconds. The verification accuracy corresponding to 50 decision trees does not reach 0.92. Therefore, the number of decision trees is determined to be 100. Each decision tree uses sampling with replacement to form a training subset from the training samples. The same sample is allowed to appear repeatedly in the training subset of the same decision tree.

[0106] In this embodiment, when each decision tree splits a node, it first selects candidate features from the core feature vector, and then calculates the splitting effect of each candidate feature.

[0107] Specifically, the number of candidate features is determined by the dimension of the core feature vector. The processing unit first takes the square root of the dimension of the core feature vector, and then takes an integer not greater than the square root result as the number of candidate features for each node. For classification tasks with state labels as normal state, observation state, abnormal evolution state, and dangerous state, the node split uses the Gini index. The calculation process of the Gini index is as follows: count the number of samples for each state label in the current node, divide the number of samples for each state label by the total number of samples in the current node to obtain the proportion of that state label; calculate the square value of each proportion, add all the square values ​​together, and then subtract the sum from 1 to obtain the Gini index of the current node.

[0108] Furthermore, after determining the candidate feature, the processing unit sequentially traverses each candidate segmentation point of the candidate feature, and each candidate segmentation point divides the current node sample into a left child node and a right child node.

[0109] The processing unit calculates the Gini index of the left and right child nodes respectively. Then, it multiplies the ratio of the number of samples in the left child node to the number of samples in the current node by the Gini index of the left child node, and multiplies the ratio of the number of samples in the right child node to the number of samples in the current node by the Gini index of the right child node. Finally, the two products are added together to obtain the split Gini index corresponding to the candidate split point. The processing unit selects the candidate feature and candidate split point with the smallest split Gini index value as the split condition of the current node. The node splitting continues until the tree depth reaches 12 layers, or the number of samples in the node reaches 10, or all samples in the node belong to the same state label.

[0110] Furthermore, after the random forest model has been trained, the processing unit calculates the contribution of each core feature vector to node splitting.

[0111] Whenever a core feature vector is used as a node splitting condition, the processing unit calculates the difference between the Gini index of the node before splitting and the Gini index after splitting, and records the difference as the contribution value of the core feature vector in that node. After training, all the contribution values ​​generated by the same core feature vector in 100 decision trees are added together to obtain the total contribution value of the core feature vector.

[0112] Subsequently, the total contribution values ​​of each core feature vector were sorted from largest to smallest, and the top 10 core feature vectors were selected as the sources of interaction features.

[0113] Specifically, determining the interaction features requires tracing back to the original feature composition corresponding to the core feature vector. The processing unit reads the loading values ​​of each original feature in the core feature vector from the principal component analysis algorithm. The original features with an absolute value of 0.3 are identified as the constituent features of the core feature vector.

[0114] The load threshold of 0.3 was determined through historical sample verification: the candidate load thresholds were set to 0.2, 0.3, and 0.4. The verification and identification accuracy corresponding to the load threshold of 0.3 reached 0.92, and the number of interaction features was 10; the number of interaction features corresponding to the load threshold of 0.2 exceeded 10, increasing the burden of manual review; the verification and identification accuracy corresponding to the load threshold of 0.4 did not reach 0.92, so the load threshold was determined to be 0.3. The processing unit obtained the interaction features between the cargo type, physicochemical properties, and climatic conditions of the target dry bulk cargo based on the load backtracking results. The interaction features included the interaction features between cargo type and moisture content, the interaction features between moisture absorption rate and relative humidity in the warehouse, the interaction features between moisture absorption and heat release coefficient and temperature change rate, the interaction features between porosity and ventilation speed, and the interaction features between critical moisture content and dew point difference.

[0115] The processing unit can be a computing device equipped with a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, the processing unit executes the intelligent identification and low false alarm graded early warning method for abnormal states of dry bulk cargo storage described in this application.

[0116] Example 3:

[0117] S2 includes obtaining an environmental fluctuation data matrix based on interaction characteristics; decomposing the environmental fluctuation data matrix to obtain multi-scale environmental decomposition components; reducing the dimensionality of the multi-scale environmental decomposition components to obtain core environmental feature vectors, and obtaining hygroscopic and exothermic physical properties to construct boundary constraints; constructing an input tensor based on the core environmental feature vectors and boundary constraints, and processing the input tensor through a long short-term memory network to obtain a cargo state sequence.

[0118] In this embodiment, the processing unit determines the content to be collected for the environmental fluctuation data matrix based on the interaction characteristics.

[0119] Specifically, the interaction characteristics include the interaction characteristics between cargo type and moisture content, the interaction characteristics between moisture absorption rate and relative humidity in the warehouse, the interaction characteristics between moisture absorption and heat release coefficient and temperature change rate, the interaction characteristics between porosity and ventilation speed, and the interaction characteristics between critical moisture content and dew point difference. The processing unit retrieves the temperature, relative humidity, dew point temperature, ventilation speed, cargo surface temperature, and cargo moisture content data in the corresponding warehouse based on the above interaction characteristics.

[0120] The sampling period was set to 5 minutes, the continuous collection time was set to 72 hours, and 4 measuring points were set in the same warehouse. The 4 measuring points were respectively located on the surface of the cargo stack, the middle layer of the cargo stack, the bottom layer of the cargo stack, and near the ventilation opening.

[0121] Specifically, the environmental fluctuation data matrix is ​​established according to the sampling time and environmental variables. Each row corresponds to one sampling time, and each column corresponds to one environmental variable or a fluctuation variable calculated from the environmental variable. The items include temperature, relative humidity, dew point temperature, ventilation velocity, cargo surface temperature, moisture content, temperature change rate, relative humidity change rate, dew point difference, moisture content change, and ventilation velocity change.

[0122] The temperature change rate is obtained by subtracting the temperature of the previous sampling time from the temperature at the current sampling time, and then dividing by the time interval between the two sampling times; the relative humidity change rate and the ventilation speed change are calculated based on the difference between two adjacent sampling times; the dew point difference is obtained by subtracting the surface temperature of the goods from the dew point temperature; and the moisture content change is obtained by subtracting the moisture content of the previous sampling time from the moisture content at the current sampling time. After the above calculations are completed, the processing unit writes the data generated by each measuring point at the same sampling time into the matrix in sequence according to the measuring point number.

[0123] Specifically, before the matrix enters the decomposition process, the processing unit performs validity processing on the sampled data.

[0124] For data with a continuous missing duration of no more than 15 minutes, the first and last valid values ​​before and after the missing point are read, the two valid values ​​are added together and divided by 2, and this sum is used as the missing point completion value. For data with a continuous missing duration of more than 15 minutes, the corresponding time period is marked as an invalid time period and removed from the state sequence calculation sample. For mutation data, the processing unit compares the difference between two adjacent sampling times. When the temperature difference exceeds 5°C, the relative humidity difference exceeds 15%, or the ventilation speed difference exceeds 1.5 meters per second, the two valid sampling values ​​before and after the sampling point are read, the four valid sampling values ​​are added together and divided by 4, and the resulting value is used to replace the mutation data.

[0125] Furthermore, when decomposing the environmental fluctuation data matrix, the processing unit uses three time scales to generate multi-scale environmental decomposition components.

[0126] Specifically, the first scale window contains 6 consecutive sampling points, corresponding to 30 minutes of data; the second scale window contains 72 consecutive sampling points, corresponding to 6 hours of data; and the third scale window contains 864 consecutive sampling points, corresponding to 72 hours of data. For each column of environmental fluctuation data, the processing unit calculates the window average, window maximum, window minimum, window change amplitude, and window trend slope at the three time scales. The window average is obtained by summing all the sampled values ​​within the window and dividing by the number of sampled values. The window change amplitude is obtained by subtracting the window minimum from the window maximum. The window trend slope is obtained through linear fitting. Specifically, the average value of each sampling time and the average value of each sampled value within the window are calculated first. Then, the difference between each sampling time and the average value of each sampling time, and the difference between each sampled value and the average value of each sampled value are multiplied and summed. Finally, the sum of the squares of the differences between each sampling time and the average value of each sampling time is divided.

[0127] The resulting 30-minute, 6-hour, and 72-hour decomposition components constitute the multi-scale environmental decomposition components.

[0128] Furthermore, after the multi-scale environment decomposition components are formed, the processing unit determines whether the number of feature columns exceeds the dimensionality reduction trigger value. The dimensionality reduction trigger value is set to 40, which is determined by historical verification: warehouse samples that have completed status labeling within the last 12 months are used for verification with 30, 40, and 50 candidate feature column numbers, respectively; when the candidate value is 30, the status recognition accuracy is 0.89; when the candidate value is 40, the status recognition accuracy is 0.93, and the average processing time per sample is 180 milliseconds; when the candidate value is 50, the status recognition accuracy is 0.934, and the average processing time per sample is 260 milliseconds.

[0129] Since the accuracy increment of candidate value 50 compared to candidate value 40 is 0.004 and the processing time increases by 80 milliseconds, the processing unit selects 40 as the dimensionality reduction trigger value.

[0130] Furthermore, when the number of feature columns of the multi-scale environmental decomposition components exceeds 40, the processing unit uses the principal component analysis algorithm to generate the core environmental feature vector.

[0131] Specifically, during processing, each column of decomposed components is first standardized. This involves calculating the mean and standard deviation of all sample values ​​in that column, subtracting the mean from each sample value, and then dividing by the standard deviation to obtain the standardized eigenvalue. Subsequently, the covariance between each eigenvalue column is calculated. For any two eigenvalue columns, the two standardized eigenvalues ​​from the same sample record are multiplied together, and the product of all sample records is summed and then divided by the number of sample records minus 1 to obtain the covariance between the two eigenvalue columns. All covariances are arranged in the order of the eigenvalue columns to form a covariance matrix.

[0132] As a preferred embodiment, the processing unit calculates the eigenvalues ​​and eigenvectors of the covariance matrix and arranges them in descending order of eigenvalues.

[0133] The information retention ratio was set to 95%, which was obtained by verifying three candidate values: 90%, 95%, and 98%. 90% corresponds to a state recognition accuracy of 0.90, 95% corresponds to a state recognition accuracy of 0.93, and 98% corresponds to a state recognition accuracy of 0.936. The average processing time for a single sample reached 270 milliseconds, so 95% was adopted. The processing unit started accumulating from the first ranked feature value, and divided the sum of the accumulated feature values ​​by the sum of all feature values. When the obtained ratio reached 95%, the selection of subsequent feature vectors was stopped.

[0134] Subsequently, the standardized multi-scale environmental decomposition components are multiplied with the selected feature vectors to obtain the core environmental feature vectors.

[0135] Specifically, the hygroscopic and heat-releasing physical properties are obtained from the warehousing inspection report, on-site re-inspection records, and historical inspection records of the same batch, including the hygroscopic rate, hygroscopic and heat-releasing coefficient, critical moisture content, thermal conductivity, specific heat capacity, bulk density, and porosity.

[0136] The processing unit constructs boundary constraints based on the aforementioned physical properties. The calculation basis for the moisture absorption boundary is the relative humidity and dew point difference within the warehouse: when the relative humidity within the warehouse reaches 85%, and the difference between the dew point temperature and the surface temperature of the goods reaches or exceeds 0℃, the moisture absorption boundary condition is established. The process for determining 85% is as follows: read the moisture content re-inspection records of the same type of goods from the past 12 months, and take records where the moisture content increases by more than 0.3% within 1 hour as moisture absorption samples, and check the relative humidity corresponding to each moisture absorption sample; when the relative humidity threshold is 85%, the accuracy rate of moisture absorption sample identification reaches 0.91, therefore 85% is taken as the moisture absorption boundary threshold.

[0137] Specifically, the heat release boundary is determined by the change in moisture content, the hygroscopic heat release coefficient, the specific heat capacity, and the ventilation velocity.

[0138] The processing unit multiplies the change in moisture content by the hygroscopic heat release coefficient to obtain the hygroscopic heat release per unit mass of goods. This hygroscopic heat release is then divided by the specific heat capacity to obtain the theoretical temperature rise. When the theoretical temperature rise reaches 0.5℃ and the ventilation velocity is less than 0.5 m / s, the heat release boundary condition is met. The 0.5℃ threshold is determined by historical treatment samples: historical samples that have undergone ventilation, dehumidification, or turning are read, and the theoretical temperature rise within 6 hours before the anomaly occurs is calculated. When the theoretical temperature rise threshold is 0.5℃, the accuracy rate for identifying treated samples reaches 0.92, and the false identification rate for untreated samples is 0.08. The ventilation velocity threshold of 0.5 m / s is determined by the ventilation equipment operation records. When the velocity is less than this value, the accuracy rate for identifying continuous temperature rise records in historical samples is 0.88.

[0139] Specifically, the moisture content boundary is determined by comparing the real-time moisture content with the critical moisture content.

[0140] Among them, the critical moisture content is preferentially adopted from the value in the warehousing inspection report; when there are more than 3 valid re-inspection records for the same type of goods, all valid critical moisture content test values ​​are added together and then divided by the number of valid test values. The result is used as the critical moisture content under the current storage conditions. When the real-time moisture content reaches the critical moisture content, the moisture content boundary condition is met. The processing unit converts the hygroscopic boundary, the exothermic boundary and the moisture content boundary into numerical constraint terms respectively. When the condition is met, it is recorded as 1, and when the condition is not met, it is recorded as 0.

[0141] Furthermore, when constructing the input tensor based on the core environmental feature vector and boundary constraints, the processing unit uses 24 consecutive sampling moments as one time segment, corresponding to 2 hours of data.

[0142] Specifically, within each time segment, the core environmental feature vector is written in the order of sampling time, and the moisture absorption boundary value, heat release boundary value, and moisture content boundary value are added at each sampling time. The time dimension of the input tensor corresponds to 24 sampling times, the feature dimension corresponds to the core environmental feature vector and 3 boundary constraint terms, and the measurement point dimension corresponds to 4 measurement points. This input tensor carries both environmental fluctuation features and moisture absorption and heat release boundary states, and serves as the input to the long short-term memory network.

[0143] Furthermore, after reading the input tensor, the Long Short-Term Memory (LSTM) network processes each time segment according to the sampling time sequence.

[0144] Specifically, the network is configured with two long short-term memory layers, each with 64 memory units. The 64 memory units are determined by three candidate values: 32, 64, and 128. The state recognition accuracy is 0.90 for 32 memory units, 0.93 for 64 memory units with an average processing time of 190 milliseconds per sample, and 0.935 for 128 memory units with an average processing time of 310 milliseconds per sample. Therefore, 64 memory units are used. The network reads the core environmental feature vector and boundary constraint terms at the current sampling time, while retaining the hidden state at the previous sampling time. This allows changes in temperature and humidity, moisture content, and boundary constraints to participate in the state calculation in chronological order.

[0145] In this embodiment, the output layer generates a status score and a status label at each sampling time, and the two are arranged according to the sampling time to form a cargo status sequence.

[0146] Specifically, the status score ranges from 0 to 1. The status label threshold is determined by the labeled samples within the last 12 months. The processing unit reads the status scores corresponding to normal records, review records, handling records, and emergency records respectively, and determines the threshold based on minimizing the number of misidentifications between adjacent status categories. When the status score is 0 to 0.30, a normal status is output; when the status score is greater than 0.30 and less than or equal to 0.50, an observation status is output; when the status score is greater than 0.50 and less than or equal to 0.75, an abnormal evolution status is output; and when the status score is greater than 0.75 and less than or equal to 1, a dangerous status is output.

[0147] Example 4:

[0148] S3 includes acquiring the cargo state sequence, calculating the trajectory similarity between the cargo state sequence and historical experience data; determining the target historical experience data based on the trajectory similarity, comparing the cargo state sequence with the target historical experience data to obtain the state fluctuation amplitude; if the experience matching degree corresponding to the state fluctuation amplitude is lower than the threshold, generating an abnormal state node; calculating the boundary deviation value based on the abnormal state node, and determining the initial safety boundary based on the boundary deviation value.

[0149] In this embodiment, after the processing unit obtains the cargo status sequence output by S2, it first arranges the status sequence according to the sampling time.

[0150] Specifically, each record in the cargo status sequence includes sampling time, warehouse number, measurement point number, status score, and status label. The status score ranges from 0 to 1, and the status labels include normal status, observation status, abnormal evolution status, and dangerous status. The processing unit uses 24 consecutive sampling points as one current trajectory segment. When the sampling period is 5 minutes, one current trajectory segment corresponds to a 2-hour cargo status change process. If the current cargo status sequence has fewer than 24 sampling points, trajectory similarity calculation is not performed temporarily, and sampling continues until 24 sampling points are reached.

[0151] Furthermore, historical experience data comes from historical warehousing records, inspection records, anomaly handling records, and historical early warning records. Before calculating trajectory similarity, the processing unit first filters the historical experience data according to the type of goods, warehouse structure, storage type, and inbound season.

[0152] For example, if the target dry bulk cargo is coal stored in a closed flat warehouse, the processing unit only uses historical experience data on coal stored in the same closed flat warehouse for comparison. After screening, the processing unit organizes the historical experience data into a historical status score sequence according to a 5-minute sampling interval, and divides it into historical trajectory segments by 24 consecutive sampling points. Each historical trajectory segment includes a historical status score, a historical status label, an inspection conclusion, and a disposal result.

[0153] Specifically, trajectory similarity is determined by score difference, trend difference, and label difference. When calculating score difference, the processing unit reads the status scores of the current trajectory segment and the historical trajectory segment under the same sampling number in sequence. The absolute value of the difference between the current status score and the historical status score is taken to obtain the score difference for that sampling number. Then, the 24 score differences are added together and divided by 24 to obtain the average score difference.

[0154] When calculating the trend difference, the processing unit calculates the slope of the state score change for the current trajectory segment and the historical trajectory segment respectively. The slope of the state score change is obtained as follows: first, the average of 24 sampling numbers is calculated, then the average of 24 state scores is calculated; then, the average of the sampling numbers is subtracted from each sampling number, and the average of the state scores is subtracted from each state score; then, the two differences corresponding to the same sampling point are multiplied and summed; finally, the summation result is divided by the summation result of the squares of the differences of the 24 sampling numbers to obtain the slope of the state score change. The processing unit subtracts the slope of the state score change of the historical trajectory segment from the slope of the state score change of the current trajectory segment and takes the absolute value to obtain the trend difference. When calculating the label difference, the processing unit compares the current state label and the historical state label of each of the 24 sampling points, counts the number of sampling points with inconsistent labels, and divides the number by 24 to obtain the label difference.

[0155] Specifically, the processing unit converts the average score difference, trend difference, and label difference into trajectory similarity.

[0156] In the conversion process, the average score difference is first subtracted from 1 to obtain the score similarity value. Then, the trend difference is normalized by subtracting the minimum trend difference in the training samples from the current trend difference, and then dividing by the difference between the maximum and minimum trend differences in the training samples. The trend difference normalization result is then subtracted from 1 to obtain the trend similarity value. Finally, the label difference is subtracted from 1 to obtain the label similarity value. The processing unit includes the score similarity value at a ratio of 0.50, the trend similarity value at a ratio of 0.30, and the label similarity value at a ratio of 0.20. These three results are then added together to obtain the trajectory similarity between the current trajectory segment and the historical trajectory segment. The trajectory similarity value ranges from 0 to 1.

[0157] Furthermore, when determining the target historical experience data based on trajectory similarity, the processing unit sorts the trajectory similarity between the current trajectory segment and all historical trajectory segments in descending order of value, and selects the historical trajectory segments ranked from 1st to 5th as the target historical experience data.

[0158] Specifically, the number of target historical experience data is determined to be 5. The determination process is as follows: using the labeled samples within 12 months prior to the current model training date as the verification object, the number of target historical trajectory segments is set to 3, 5, and 8 respectively for verification. When the number is 3, the initial safety boundary identification accuracy is 0.88; when the number is 5, the initial safety boundary identification accuracy is 0.92, and the average processing time per sample is 170 milliseconds; when the number is 8, the initial safety boundary identification accuracy is 0.923, and the average processing time per sample is 240 milliseconds. Since the accuracy increment of 8 compared to 5 is 0.003, and the average processing time increases by 70 milliseconds, 5 is used as the number of target historical experience data for the processing unit.

[0159] Furthermore, when comparing the cargo status sequence with historical experience data of the target, the processing unit calculates the status fluctuation amplitude.

[0160] Specifically, for each sampling point in the current trajectory segment, the processing unit reads the historical state scores of 5 target historical trajectory segments under the same sampling sequence number, adds the 5 historical state scores and divides by 5 to obtain the historical reference score of the sampling point. Then, the current state score is subtracted from the historical reference score and the absolute value is taken to obtain the state deviation value of the sampling point. The processing unit adds the state deviation values ​​of 24 sampling points and divides by 24 to obtain the average state fluctuation amplitude. At the same time, the state deviation value with the largest value among the 24 state deviation values ​​is selected as the peak state fluctuation amplitude. The average state fluctuation amplitude represents the overall deviation of the current trajectory segment relative to the target historical experience data, and the peak state fluctuation amplitude represents the deviation peak of a single sampling point.

[0161] In this embodiment, the empirical matching degree is calculated from the average fluctuation amplitude of the state and the peak fluctuation amplitude of the state.

[0162] Specifically, the processing unit first subtracts the average fluctuation amplitude of the state from 1 to obtain the overall matching value; then, it subtracts the peak fluctuation amplitude of the state from 1 to obtain the peak matching value. Subsequently, the overall matching value is included at a ratio of 0.70, and the peak matching value is included at a ratio of 0.30. The two included results are then added together to obtain the empirical matching degree. The empirical matching degree threshold is determined to be 0.80. This threshold is obtained through verification using labeled samples. The processing unit uses the inspection conclusions and handling records within the previous 12 months from the current model training date as the labeling basis, and uses 0.75, 0.80, and 0.85 as candidate thresholds, respectively. When the candidate threshold is 0.75, the accuracy rate for identifying abnormal state nodes is 0.86; when the candidate threshold is 0.80, the accuracy rate for identifying abnormal state nodes is 0.91, and the false recognition rate is 0.07; when the candidate threshold is 0.85, the false recognition rate is 0.12. Therefore, the empirical matching degree threshold is 0.80.

[0163] Furthermore, when the empirical matching degree corresponding to the state fluctuation amplitude is less than 0.80, the processing unit generates an abnormal state node.

[0164] Specifically, the recorded content of the abnormal state node includes sampling time, warehouse number, measurement point number, current status score, historical reference score, status deviation value, status label, and experience matching degree. If the experience matching degree of three consecutive sampling points for the same measurement point is less than 0.80, the processing unit merges the three sampling points into one abnormal state node segment and records the start time, end time, and duration of the node segment. Three consecutive sampling points correspond to 15 minutes. This continuous confirmation time is determined through historical sample verification: when the continuous confirmation time is 10 minutes, the false recognition rate is 0.11; when the continuous confirmation time is 15 minutes, the false recognition rate is 0.07; and when the continuous confirmation time is 20 minutes, the false recognition rate is 0.10. Therefore, 15 minutes is used as the formation condition for the abnormal state node segment.

[0165] As a preferred embodiment, when calculating the boundary deviation value based on the abnormal state node, the processing unit extracts the node deviation intensity, node duration, and node expansion range.

[0166] Specifically, the node deviation intensity is determined by averaging the state deviation values ​​of each sampling point within the abnormal state node segment. During calculation, all state deviation values ​​within the node segment are summed and then divided by the number of sampling points within the node segment. The node duration is obtained by multiplying the number of sampling points within the node segment by 5 minutes. The node expansion range is determined by the number of measurement points that generate abnormal state nodes within the same time period. The processing unit normalizes the node deviation intensity, node duration, and node expansion range respectively. During normalization, the current value is subtracted from the corresponding minimum value in the training samples and then divided by the difference between the corresponding maximum and minimum values ​​in the training samples.

[0167] Subsequently, the normalized result of node deviation intensity is included at a ratio of 0.50, the normalized result of node duration is included at a ratio of 0.30, and the normalized result of node expansion range is included at a ratio of 0.20. The three included results are then added together to obtain the boundary deviation value.

[0168] Specifically, the processing unit determines the initial safety boundary based on the boundary deviation value.

[0169] Specifically, when the boundary deviation value is less than or equal to 0.30, the corresponding cargo status sequence is included in the initial safety boundary; when the boundary deviation value is greater than 0.30 and less than or equal to 0.50, the corresponding cargo status sequence is marked as a boundary observation state, and the average status score of the cargo status sequence plus 0.10 is used as the upper limit correction value of the initial safety boundary; when the boundary deviation value is greater than 0.50, the corresponding cargo status sequence is marked as an abnormal sequence outside the boundary, and the cargo status sequence that did not generate an abnormal status node is used as the end of the initial safety boundary; the boundary deviation value thresholds of 0.30 and 0.50 are determined by historical labeled samples, and the processing unit verifies the boundary deviation value with the manual inspection conclusions and disposal records; when the boundary deviation value is 0.30 as the judgment value within the safety boundary, the accuracy rate of normal sample identification is 0.93; when the boundary deviation value is 0.50 as the judgment value of abnormality outside the boundary, the accuracy rate of disposal sample identification is 0.91.

[0170] Therefore, 0.30 is used to determine the state within the safety boundary, and 0.50 is used to determine the abnormal state outside the boundary.

[0171] After the above processing, the initial safety boundary is determined by trajectory similarity, state fluctuation amplitude, empirical matching degree and boundary deviation value. This initial safety boundary is used to limit the permissible state range of the target dry bulk cargo under the current storage conditions.

[0172] Example 5:

[0173] S4 includes calculating the difference between the cargo state sequence that deviates from historical experience data and the baseline state to obtain the state offset vector; obtaining the temporal evolution gradient based on the state offset vector, fusing the temporal evolution gradient with environmental factors to determine the risk evolution rate; if the risk evolution rate is greater than a threshold, constructing a trend matrix based on the risk evolution rate; and performing dimensionality reduction on the trend matrix to obtain the risk evolution characteristics corresponding to the cargo state sequence that deviates from historical experience data.

[0174] In this embodiment, the processing unit receives the cargo status sequence that deviates from historical experience data output by S3, and sorts it according to sampling time, warehouse number, and measurement point number.

[0175] Specifically, each cargo status record includes a status score and a status label. The status score ranges from 0 to 1, and the status labels include normal status, observation status, abnormal evolution status, and dangerous status. The processing unit uses the target historical experience data determined by S3 as the benchmark source. Under the same sampling sequence number and the same measurement point number, it reads the historical status scores from 5 target historical trajectories, adds the 5 historical status scores together, and divides by 5 to obtain the benchmark status score for the corresponding sampling point. For the benchmark status label, the processing unit counts the occurrence frequency of each status label in the 5 target historical trajectories and uses the status label with the highest occurrence frequency as the benchmark status label. If two status labels have the same occurrence frequency, the status label with the larger status level value is used as the benchmark status label. The normal status is recorded as 0, the observation status as 1, the abnormal evolution status as 2, and the dangerous status as 3.

[0176] Furthermore, when calculating the state offset vector, the processing unit reads the sampling points in the current cargo state sequence one by one.

[0177] For each sampling point, the processing unit subtracts the baseline state score from the current state score to obtain the score offset value; it also subtracts the state level value corresponding to the baseline state label from the state level value corresponding to the current state label to obtain the label offset value. The score offset value retains the positive and negative directions: a positive value indicates that the current state score is higher than the baseline state score, and a negative value indicates that the current state score is lower than the baseline state score. When the label offset value is positive, it indicates that the current state level has increased relative to the baseline state level. The processing unit writes the score offset value and label offset value of the same sampling point into the offset record and arranges all offset records in the order of sampling time to obtain the state offset vector.

[0178] Furthermore, after obtaining the state offset vector, the processing unit calculates the temporal evolution gradient.

[0179] Specifically, for two adjacent sampling points, the processing unit subtracts the score offset value of the previous sampling point from the score offset value of the subsequent sampling point, and then divides it by 5 minutes to obtain the score offset gradient; it subtracts the label offset value of the previous sampling point from the label offset value of the subsequent sampling point, and then divides it by 5 minutes to obtain the label offset gradient; when 24 consecutive sampling points are taken as one analysis segment, the processing unit obtains 23 score offset gradients and 23 label offset gradients.

[0180] Subsequently, the processing unit adds up the 23 rating offset gradients and divides by 23 to obtain the average rating offset gradient; it also adds up the 23 label offset gradients and divides by 23 to obtain the average label offset gradient; the average rating offset gradient and the average label offset gradient together constitute the temporal evolution gradient, which is used to characterize the process of deviation from the state increasing or decreasing over time.

[0181] Specifically, environmental factors are obtained from environmental fluctuation data corresponding to the cargo state sequence, including temperature change rate, relative humidity change rate, moisture content change, ventilation speed change, and dew point risk item.

[0182] The temperature change rate is calculated by subtracting the temperature from the previous sampling time from the current sampling time, and then dividing by 5 minutes. The relative humidity change rate is calculated by subtracting the relative humidity from the previous sampling time from the current sampling time, and then dividing by 5 minutes. The moisture content change rate is calculated by subtracting the moisture content from the previous sampling time from the current sampling time. The ventilation speed change rate is calculated by subtracting the ventilation speed from the previous sampling time from the current sampling time. The dew point risk item is determined based on the relationship between the dew point temperature and the cargo surface temperature. When the result of subtracting the cargo surface temperature from the dew point temperature reaches 0℃ or is greater than 0℃, the dew point risk item is recorded as 1; when the result is less than 0℃, the dew point risk item is recorded as 0.

[0183] Before integrating the temporal evolution gradient and environmental factors, the processing unit normalizes the average score offset gradient, average label offset gradient, temperature change rate, relative humidity change rate, moisture content change, and ventilation speed change.

[0184] Furthermore, the normalization process is completed using the minimum and maximum values ​​in the training samples. Specifically, the current value is subtracted from the minimum value of the corresponding indicator in the training samples, and then divided by the difference between the maximum and minimum values ​​of the corresponding indicator in the training samples to obtain a normalized result with a value range of 0 to 1. Since the dew point risk item is already 0 or 1, no further normalization processing is performed.

[0185] Specifically, the rate of risk evolution is determined using a weighted approach.

[0186] The processing unit incorporates the normalized results of the average score offset gradient (0.30), the average label offset gradient (0.20), the temperature change rate (0.15), the relative humidity change rate (0.15), the moisture content change (0.10), the ventilation velocity change (0.05), and the dew point risk item (0.05). These incorporations are then summed to obtain the risk evolution rate.

[0187] The weights mentioned above were determined by historical sample verification. The processing unit used the disposal records within 12 months prior to the current model training date as the verification basis and tested 3 sets of weight combinations respectively. When the total weight of the average score offset gradient and the average label offset gradient was 0.50 and the total weight of the environmental factors was 0.50, the accuracy of abnormal evolution sample identification was 0.92 and the false identification rate was 0.07. Therefore, this set of weights was adopted.

[0188] In this embodiment, the threshold for the risk evolution rate is set to 0.60. When the threshold is set, the processing unit calls up the warehouse samples that have been labeled within the previous 12 months prior to the current model training date. The labeling is based on manual inspection conclusions, ventilation records, dehumidification records, turning records, and emergency work order records.

[0189] The processing unit used 0.50, 0.60, and 0.70 as candidate thresholds for verification. When the candidate threshold was 0.50, the false recognition rate was 0.13. When the candidate threshold was 0.60, the accuracy rate for identifying abnormal evolution samples was 0.92, and the false recognition rate was 0.07. When the candidate threshold was 0.70, the missed recognition rate was 0.11. Therefore, 0.60 was determined as the risk evolution rate threshold. When the risk evolution rate was greater than 0.60, the processing unit constructed a trend matrix based on the risk evolution rate. When the risk evolution rate was less than or equal to 0.60, the processing unit marked the corresponding cargo state sequence as an observation sequence and did not proceed to the trend matrix construction process.

[0190] Furthermore, when constructing the trend matrix, the processing unit uses 24 consecutive sampling points as one trend analysis segment, corresponding to 2 hours of data. Each row in the trend matrix corresponds to one sampling point, and each column corresponds to one trend indicator.

[0191] Specifically, the trend indicators include risk evolution rate, score offset value, score offset gradient, label offset value, label offset gradient, temperature change rate, relative humidity change rate, moisture content change, ventilation speed change, dew point risk item, and number of abnormal measurement points. The number of abnormal measurement points is determined by the number of measurement points with a risk evolution rate greater than 0.60 at the same sampling time. The processing unit writes the above trend indicators in the order of sampling time so that the trend matrix reflects the increasing process of the deviation state in the time dimension and the expansion process of abnormal measurement points.

[0192] Specifically, before performing dimensionality reduction on the trend matrix, the processing unit determines whether the number of feature columns in the trend matrix exceeds 8.

[0193] The dimensionality reduction trigger value 8 was determined by historical sample verification. The processing unit used 6, 8, and 10 as candidate trigger values ​​for verification. When the trigger value was 6, the accuracy rate of risk evolution feature identification was 0.89. When the trigger value was 8, the accuracy rate of risk evolution feature identification was 0.92, and the average processing time per sample was 160 milliseconds. When the trigger value was 10, the accuracy rate of risk evolution feature identification was 0.924, and the average processing time per sample was 230 milliseconds. Since the accuracy increment of trigger value 10 relative to trigger value 8 was 0.004, and the average processing time increased by 70 milliseconds, 8 was chosen as the dimensionality reduction trigger value.

[0194] When the number of feature columns is less than or equal to 8, the processing unit directly uses the trend matrix as the source of risk evolution features; when the number of feature columns is greater than 8, the processing unit performs principal component analysis for dimensionality reduction.

[0195] Furthermore, during principal component analysis dimensionality reduction, the processing unit first standardizes each column of trend indicators in the trend matrix.

[0196] Specifically, the standardization process is as follows: calculate the mean and standard deviation of all sample values ​​in the column, subtract the mean from each sample value, and then divide by the standard deviation to obtain the standardized trend feature; subsequently, the processing unit calculates the covariance between any two columns of standardized trend features. During the calculation, the two columns of standardized trend features under the same sampling record are multiplied, the product of all sampling records is summed, and then divided by the number of sampling records minus 1 to obtain the covariance between the two columns; after all covariances are arranged in the order of the feature columns, a covariance matrix is ​​formed. The processing unit obtains the eigenvalues ​​and eigenvectors of the covariance matrix and arranges them in descending order of eigenvalues.

[0197] As a preferred embodiment, the information retention ratio is 95%, which is obtained by verifying three candidate ratios: 90%, 95%, and 98%. When the candidate ratio is 90%, the risk evolution identification accuracy is 0.90; when the candidate ratio is 95%, the risk evolution identification accuracy is 0.92, and the average processing time per sample is 170 milliseconds; when the candidate ratio is 98%, the risk evolution identification accuracy is 0.925, and the average processing time per sample is 250 milliseconds. Since the accuracy increment of 98% compared to 95% is 0.005, and the average processing time increases by 80 milliseconds, 95% is adopted. The processing unit starts accumulating from the first ranked feature value, and divides the sum of the accumulated feature values ​​by the sum of all feature values. When the obtained ratio reaches 95%, the selection of subsequent feature vectors stops.

[0198] Subsequently, the standardized trend matrix is ​​multiplied by the selected feature vectors to obtain the dimensionality-reduced trend features.

[0199] Furthermore, after the dimensionality reduction process is completed, the processing unit establishes a correspondence between the dimensionality-reduced trend features and the risk evolution rate, number of abnormal measurement points, and dew point risk items in the original trend matrix, and generates risk evolution features corresponding to cargo state sequences that deviate from historical experience data; the risk evolution features include offset growth features, state level change features, environment-driven features, dew point triggering features, and measurement point expansion features.

[0200] Example 6:

[0201] S5 includes obtaining an initial warning standard, which is obtained by comparing the risk evolution characteristics with the historical safety baseline; calculating the projection distance of the risk evolution characteristics within the corresponding interval of the initial warning standard to obtain a deviation value; obtaining a first critical value based on the deviation value; determining whether the first critical value exceeds a preset threshold; if the first critical value exceeds the preset threshold, determining the cargo state sequence corresponding to the first critical value as a first danger signal.

[0202] In this embodiment, the processing unit receives the risk evolution characteristics output by S4.

[0203] Specifically, the risk evolution characteristics include offset growth characteristics, state level change characteristics, environment-driven characteristics, dew point triggering characteristics, and measurement point expansion characteristics. The processing unit calls the historical safety baseline as the comparison object. The historical safety baseline is formed by normal storage samples within 12 months prior to the current model training date. The screening conditions for normal storage samples are: the manual inspection conclusion is normal, no ventilation treatment record is generated, no dehumidification treatment record is generated, no turning treatment record is generated, no isolation treatment record is generated, and no emergency work order is generated.

[0204] Furthermore, when establishing historical safety baselines, the processing unit reads the offset growth characteristics, state level change characteristics, environmental driving characteristics, dew point triggering characteristics, and measurement point expansion characteristics from normal storage samples.

[0205] For each type of feature, the processing unit first adds up the corresponding feature values ​​of all normal storage samples, then divides by the number of normal storage samples to obtain the safety benchmark value for that type of feature. Subsequently, the safety benchmark value is subtracted from the corresponding feature value of each normal storage sample, the difference is squared, all squared values ​​are added up and divided by the number of normal storage samples, and the square root of the result is taken to obtain the safety fluctuation value for that type of feature. The safety benchmark value and the safety fluctuation value together constitute the historical safety baseline.

[0206] Specifically, the initial warning standard is obtained by comparing the risk evolution characteristics with the historical safety baseline. For each type of risk evolution characteristic in the current cargo status sequence, the processing unit first subtracts the corresponding safety benchmark value from the current characteristic value, and then divides the difference by the corresponding safety fluctuation value to obtain the standardized deviation value.

[0207] Specifically, when the safety fluctuation value is 0, the processing unit replaces the safety fluctuation value with 0.001 before performing the division calculation to avoid calculation interruption; when the standardized deviation value is less than or equal to 1, it corresponds to the safe interval; when the standardized deviation value is greater than 1 and less than or equal to 2, it corresponds to the observation interval; when the standardized deviation value is greater than 2 and less than or equal to 3, it corresponds to the warning interval; when the standardized deviation value is greater than 3, it corresponds to the danger interval.

[0208] Specifically, the interval boundary values ​​1, 2, and 3 mentioned above were determined through verification using historical labeled samples.

[0209] The processing unit retrieves normal records, manual review records, warehousing and disposal records, and emergency work order records from the previous 12 months prior to the current model training date. The accuracy rates of manual review record recognition, warehousing and disposal record recognition, and emergency work order record recognition are used as verification indicators. The candidate boundary combinations are set to 0.8, 1.8, 2.8; 1, 2, 3; and 1.2, 2.2, 3.2, respectively. When the candidate boundary combination is 1, 2, 3, the accuracy rate of manual review record recognition is 0.90, the accuracy rate of warehousing and disposal record recognition is 0.92, and the accuracy rate of emergency work order record recognition is 0.91, with a false recognition rate of 0.07. In the other two candidate boundary combinations, at least one recognition accuracy rate does not reach 0.90.

[0210] Therefore, the processing unit uses 1, 2, and 3 as the interval boundary values ​​for the initial early warning standard.

[0211] Furthermore, when calculating the projected distance of risk evolution features within the interval corresponding to the initial warning standard, the processing unit first determines the interval in which each type of risk evolution feature is located, and then calculates the distance of the feature value relative to the lower limit of the interval within its respective interval. For the safe interval, the lower limit is 0 and the upper limit is 1; for the observation interval, the lower limit is 1 and the upper limit is 2; for the warning interval, the lower limit is 2 and the upper limit is 3. The processing unit subtracts the lower limit of the interval from the current standardized deviation value, and then divides it by the difference between the upper limit and the lower limit of the interval to obtain the projected distance of the feature within its respective interval.

[0212] For the danger zone, the lower limit is 3, and the upper limit is determined by the maximum standardized deviation value in the historical emergency work order samples; when the current standardized deviation value exceeds the upper limit of the zone, the projected distance is recorded as 1.

[0213] Furthermore, after obtaining the projected distances of the five types of risk evolution characteristics, the processing unit calculates the deviation values.

[0214] Specifically, the projected distance of the offset growth feature is calculated as 0.30, the projected distance of the state level change feature is calculated as 0.20, the projected distance of the environment-driven feature is calculated as 0.20, the projected distance of the dew point trigger feature is calculated as 0.10, and the projected distance of the measurement point expansion feature is calculated as 0.20. The processing unit adds the above five calculated results to obtain the deviation value. The above weights are determined by verification using labeled samples within the 12 months prior to the current model training date. During verification, the processing unit compares three sets of weights respectively. When the total weight of the offset growth feature and the measurement point expansion feature is 0.50, the total weight of the state level change feature and the environment-driven feature is 0.40, and the weight of the dew point trigger feature is 0.10, the accuracy of the first danger signal identification is 0.92, and the false identification rate is 0.07. Therefore, this set of weights is adopted.

[0215] Specifically, the first critical value is obtained based on both the deviation value and the interval level.

[0216] The processing unit first determines an interval level value for each type of risk evolution characteristic: 0 for the safe interval, 0.30 for the observation interval, 0.60 for the warning interval, and 0.85 for the danger interval. Then, the processing unit reads the interval level values ​​corresponding to the five types of risk evolution characteristics and selects the interval level value with the largest value as the interval base value. The processing unit adds the deviation value (0.60) and the interval base value (0.40), and adds the two values ​​together to obtain the first critical value. If the first critical value is greater than 1, the processing unit corrects it to 1; if the first critical value is less than 0, the processing unit corrects it to 0.

[0217] Specifically, the preset threshold is set to 0.60.

[0218] The processing unit calls upon labeled warehouse samples from the previous 12 months prior to the current model training date, using manual inspection conclusions, ventilation records, dehumidification records, turning records, isolation records, and emergency work orders as verification criteria. Candidate thresholds are set to 0.50, 0.60, and 0.70, respectively. At a candidate threshold of 0.50, the false recognition rate is 0.12; at a candidate threshold of 0.60, the accuracy rate for identifying the first danger signal is 0.92, with a false recognition rate of 0.07; and at a candidate threshold of 0.70, the missed recognition rate is 0.10.

[0219] Therefore, the processing unit determines 0.60 as the preset threshold for the first critical value.

[0220] In this embodiment, when determining whether the first critical value exceeds the preset threshold, the processing unit compares the first critical value with 0.60.

[0221] Specifically, when the first critical value is greater than 0.60, the processing unit determines that the first critical value exceeds a preset threshold and marks the cargo status sequence corresponding to the first critical value as the first danger signal; when the first critical value is less than or equal to 0.60, the processing unit does not generate the first danger signal and retains the corresponding cargo status sequence as an observation sequence; the recorded content of the first danger signal includes the warehouse number, measurement point number, sampling time, risk evolution characteristics, deviation value, first critical value, the magnitude of exceeding the threshold, and the corresponding cargo status sequence; wherein, the magnitude of exceeding the threshold is obtained by subtracting 0.60 from the first critical value.

[0222] Example 7:

[0223] S6 includes obtaining the initial risk level corresponding to the first danger signal, obtaining real-time climate conditions based on the initial risk level, performing dimensionality reduction processing on the real-time climate conditions to obtain core meteorological data, determining the fusion weight based on the core meteorological data, fusing the first danger signal and the core meteorological data through the fusion weight to obtain a second critical value, determining whether the second critical value is greater than the first critical value, and if the second critical value is greater than the first critical value, extracting the second danger signal corresponding to the second critical value.

[0224] In this embodiment, the processing unit receives the first danger signal generated by S5 and reads the warehouse number, measuring point number, sampling time, first critical value, deviation value and risk evolution characteristics in the first danger signal.

[0225] Specifically, the processing unit determines the initial risk level based on the first critical value: when the first critical value is greater than 0.60 and less than or equal to 0.75, it is determined to be level 1; when the first critical value is greater than 0.75 and less than or equal to 0.90, it is determined to be level 2; when the first critical value is greater than 0.90 and less than or equal to 1, it is determined to be level 3. The above-mentioned boundary values ​​are consistent with the determination range of the first critical value in S5, and are used to limit the data call range of real-time climate conditions.

[0226] Furthermore, when obtaining real-time climate conditions based on the initial risk level, the processing unit calls the sensor records according to the warehouse number and measuring point number corresponding to the first hazard signal; when the initial risk level is level 1, data from the 2 hours prior to the sampling time is called; when the initial risk level is level 2, data from the 4 hours prior to the sampling time is called; when the initial risk level is level 3, data from the 6 hours prior to the sampling time is called; the sampling cycle is 5 minutes, therefore 2 hours corresponds to 24 sampling points, 4 hours to 48 sampling points, and 6 hours to 72 sampling points; real-time climate conditions include warehouse temperature, warehouse relative humidity, dew point temperature, ventilation speed, temperature difference between inside and outside the warehouse, air pressure, and ventilation equipment operating status.

[0227] Specifically, the processing unit performs data verification after obtaining real-time climate conditions.

[0228] When a single sampling point is missing, the preceding and following valid values ​​of that sampling point are read, the two valid values ​​are added together and divided by 2 to obtain the filler value. When three consecutive sampling points are missing, the data of that sampling point within the corresponding time window are removed, and the data of the remaining valid sampling points in the same compartment are used for calculation. For abrupt changes, the processing unit compares the difference between two adjacent sampling points. When the temperature difference exceeds 5°C, the relative humidity difference exceeds 15%, or the ventilation speed difference exceeds 1.5 m / s, the two valid sampling values ​​before and after that sampling point are read, the four valid sampling values ​​are added together and divided by 4, and the resulting value is used to replace the abrupt change data.

[0229] In this embodiment, before performing dimensionality reduction processing on real-time climate conditions, the processing unit constructs a real-time climate feature table.

[0230] Specifically, the feature table is arranged according to the sampling time, with each row corresponding to one sampling time and each column corresponding to one climate feature. The climate features include the temperature inside the warehouse, the relative humidity inside the warehouse, the dew point temperature, the ventilation speed, the temperature difference between inside and outside the warehouse, the air pressure value, the rate of temperature change, the rate of humidity change, the dew point difference, and the change in wind speed. The rate of temperature change is obtained by subtracting the temperature inside the warehouse from the temperature at the previous sampling time from the temperature at the current sampling time, and then dividing by 5 minutes. The rate of humidity change is obtained by subtracting the relative humidity at the previous sampling time from the relative humidity at the current sampling time, and then dividing by 5 minutes. The dew point difference is obtained by subtracting the surface temperature of the goods from the dew point temperature. The change in wind speed is obtained by subtracting the ventilation speed at the previous sampling time from the ventilation speed at the current sampling time.

[0231] As a preferred implementation, the dimensionality reduction process employs principal component analysis (PCA) algorithm.

[0232] Specifically, the processing unit first standardizes each column in the real-time climate feature table by calculating the mean and standard deviation of all sampled values ​​in that column, subtracting the mean from each sampled value and then dividing by the standard deviation to obtain the standardized climate feature. Subsequently, the covariance between any two columns of standardized climate features is calculated by multiplying the two columns of standardized climate features at the same sampling time, summing the products over all sampling times, and then dividing by the number of sampling points minus 1. All covariances are arranged in the order of the feature columns to form a covariance matrix. The processing unit then obtains the eigenvalues ​​and eigenvectors of the covariance matrix and arranges them in descending order of eigenvalues.

[0233] Specifically, the information retention ratio is determined to be 95%, which is obtained by verifying three candidate ratios: 90%, 95%, and 98%. When the candidate ratio is 90%, the accuracy rate of identifying the second danger signal is 0.90; when the candidate ratio is 95%, the accuracy rate of identifying the second danger signal is 0.93, and the average processing time per sample is 170 milliseconds; when the candidate ratio is 98%, the accuracy rate of identifying the second danger signal is 0.935, and the average processing time per sample is 260 milliseconds. Since the accuracy increment of 98% compared to 95% is 0.005, and the average processing time increases by 90 milliseconds, the processing unit adopts 95%. The processing unit starts accumulating from the first ranked feature value, and divides the sum of the accumulated feature values ​​by the sum of all feature values. When the obtained ratio reaches 95%, the selection of subsequent feature vectors stops.

[0234] Subsequently, the standardized climate characteristics are multiplied by the selected feature vectors to obtain the core meteorological data.

[0235] In this embodiment, core meteorological data is used to determine fusion weights, and the processing unit extracts humidity-driven terms, temperature-driven terms, dew point-driven terms, and ventilation inhibition terms from the core meteorological data.

[0236] Among them, the humidity driving term is determined by the load results corresponding to the relative humidity and humidity change rate in the reservoir; the temperature driving term is determined by the load results corresponding to the temperature and temperature change rate in the reservoir; the dew point driving term is determined by the load results corresponding to the dew point difference; the ventilation suppression term is determined by the load results corresponding to the ventilation wind speed and wind speed change; climatic characteristics with an absolute load value of 0.30 are included in the corresponding driving term; the load threshold of 0.30 is obtained by verification from three candidate values ​​of 0.20, 0.30 and 0.40; when the candidate value is 0.30, the accuracy of the second danger signal identification reaches 0.93, and the number of core meteorological data items is 4, so 0.30 is adopted as the load threshold.

[0237] As a preferred embodiment, the processing unit normalizes the humidity-driven, temperature-driven, dew-point-driven, and ventilation-suppression terms respectively.

[0238] Specifically, during normalization, the current driving term value is subtracted from the corresponding minimum value in the training samples, and then divided by the difference between the corresponding maximum and minimum values ​​in the training samples to obtain a driving term value ranging from 0 to 1. The driving term trigger value is determined to be 0.70, which is obtained by verifying three candidate values: 0.60, 0.70, and 0.80. When the candidate value is 0.60, the false recognition rate is 0.11; when the candidate value is 0.70, the accuracy rate of the second danger signal recognition is 0.93, and the false recognition rate is 0.06; when the candidate value is 0.80, the missed recognition rate is 0.10. Therefore, the processing unit uses 0.70 as the driving term trigger value.

[0239] Specifically, the fusion weights are determined based on the triggering results of the driving terms.

[0240] Specifically, when the humidity driving item value reaches 0.70, the humidity weight is recorded as 0.30; when the temperature driving item value reaches 0.70, the temperature weight is recorded as 0.25; when the dew point driving item value reaches 0.70, the dew point weight is recorded as 0.25; when the ventilation suppression item value reaches 0.70, the ventilation weight is recorded as 0.20; driving items that do not reach 0.70 have a corresponding weight of 0; when there are untriggered driving items, the processing unit adds up the weights of all triggered items to obtain a total weight, and then divides each triggered weight by the total weight to make the sum of the weights participating in the fusion equal to 1.

[0241] Furthermore, when fusing the first danger signal with the core meteorological data through fusion weights, the processing unit first calculates the meteorological correction value.

[0242] Specifically, the calculation process is as follows: multiply the humidity-driven item value by the humidity weight, multiply the temperature-driven item value by the temperature weight, multiply the dew point-driven item value by the dew point weight, multiply the ventilation suppression item value by the ventilation weight, and then add the resulting products to obtain the meteorological correction value; subsequently, the processing unit includes the first critical value at 0.70, includes the meteorological correction value at 0.30, and adds the two included results to obtain the second critical value; if the second critical value is greater than 1, it is corrected to 1; if the second critical value is less than 0, it is corrected to 0.

[0243] Specifically, the fusion ratio of the first critical value and the meteorological correction value is determined by historical sample verification.

[0244] The processing unit was verified using three candidate ratios: 0.80 and 0.20, 0.70 and 0.30, and 0.60 and 0.40. When the first critical value was 0.70 and the meteorological correction value was 0.30, the accuracy of the second danger signal identification was 0.93 and the false identification rate was 0.06. Therefore, 0.70 and 0.30 were used as the fusion ratios.

[0245] Furthermore, when determining whether the second critical value is greater than the first critical value, the processing unit directly compares the two values. If the second critical value is greater than the first critical value, it indicates that the risk of the real-time climate conditions forming the first danger signal has increased, and the processing unit extracts the second danger signal corresponding to the second critical value. If the second critical value is less than or equal to the first critical value, no second danger signal is generated, and the first danger signal is retained as a signal to be observed. The second danger signal includes the warehouse number, measuring point number, sampling time, first critical value, second critical value, core meteorological data, fusion weight, meteorological correction value, and corresponding cargo status sequence.

[0246] Example 8:

[0247] S7 includes acquiring a second danger signal and extracting a feature sequence, processing the feature sequence within a preset time window to obtain the signal strength and fluctuation frequency; calculating the attenuation amplitude based on the signal strength and fluctuation frequency, and determining the duration of the second danger signal based on the attenuation amplitude; determining whether the duration meets the preset confirmation conditions; if the duration meets the preset confirmation conditions, retaining the second danger signal; if the duration does not meet the preset confirmation conditions, filtering or downgrading the second danger signal.

[0248] In this embodiment, after the processing unit obtains the second danger signal output by S6, it reads the warehouse number, measuring point number, generation time, second critical value, core meteorological data, meteorological correction value and corresponding cargo status sequence from the signal.

[0249] Specifically, the processing unit uses the generation time of the second danger signal as the starting point for continuous confirmation and collects data for 2 hours thereafter as a preset time window. When the sampling period is 5 minutes, 24 sampling points are formed within the preset time window. The preset time window of 2 hours is determined through historical verification: the verification samples are the labeled records within 12 months before the current model training date. When the window is 1 hour, the instantaneous disturbance false recognition rate is 0.10. When the window is 2 hours, the second danger signal confirmation accuracy rate is 0.93 and the false recognition rate is 0.06. When the window is 3 hours, the proportion of records with delayed confirmation results is 0.12. Therefore, 2 hours is used as the preset time window.

[0250] In this embodiment, the processing unit extracts the feature sequence of the second danger signal within a preset time window.

[0251] Specifically, the feature sequences include a second critical value sequence, a state score sequence, a risk evolution rate sequence, a meteorological correction value sequence, an abnormal measurement point number sequence, and a dew point risk item sequence; the second critical value sequence is output point by point by S6; the state score sequence is obtained from the cargo state sequence generated by S2; the risk evolution rate sequence is obtained by S4; the meteorological correction value sequence is calculated point by point by core meteorological data and fusion weights; the abnormal measurement point number sequence is obtained by counting the number of measurement points that generate the second danger signal at the same sampling time; the dew point risk item is obtained by subtracting the cargo surface temperature from the dew point temperature, and is recorded as 1 when the difference reaches 0℃ or is greater than 0℃, and as 0 when the difference is less than 0℃.

[0252] Specifically, before calculating the signal strength, the processing unit performs numerical processing on the feature sequence.

[0253] Among them, the values ​​of the second critical value, state score, risk evolution rate, meteorological correction value and dew point risk item range from 0 to 1; the number of abnormal measurement points is converted according to the total number of measurement points in the warehouse, specifically by dividing the number of abnormal measurement points at the current sampling time by the total number of measurement points in the warehouse to obtain the abnormal measurement point ratio; when there are 4 measurement points in the warehouse and there are 2 abnormal measurement points at the current sampling time, the abnormal measurement point ratio is 0.5.

[0254] Specifically, the signal strength is calculated point by point.

[0255] The processing unit calculates the second critical value as 0.35, the state score as 0.20, the risk evolution rate as 0.20, the meteorological correction value as 0.10, the proportion of abnormal measuring points as 0.10, and the dew point risk item as 0.05. The results of the six calculations are added together to obtain the signal strength of the sampling point.

[0256] The aforementioned weights were determined by validation using labeled samples within the 12 months prior to the current model training date. When the combined weight of the second critical value, state score, and risk evolution rate was 0.75, and the combined weight of the meteorological correction value, the proportion of abnormal measurement points, and the dew point risk item was 0.25, the accuracy rate for identifying persistent danger signals was 0.93, and the false recognition rate was 0.06. Therefore, this set of weights was adopted. The processing unit arranged the signal strengths of the 24 sampling points according to the sampling time to form a signal strength sequence; then, the 24 signal strengths were added together and divided by 24 to obtain the window average signal strength.

[0257] As a preferred embodiment, the fluctuation frequency is determined by the number of adjacent changes in the signal strength sequence.

[0258] Specifically, the processing unit sequentially reads the signal strength of two adjacent sampling points, subtracts the signal strength of the previous sampling point from the signal strength of the subsequent sampling point, and takes the absolute value of the difference. When the absolute value reaches 0.08, it is recorded as one valid fluctuation. When the absolute value is less than 0.08, it is not counted as a valid fluctuation. The valid fluctuation judgment value of 0.08 is determined through normal storage samples. The processing unit reads normal storage samples that have not generated disposal records within 12 months before the current model training date and counts the signal strength difference between adjacent sampling points. When the judgment value is 0.08, the instantaneous noise in the normal storage samples is eliminated, and the accuracy rate of identifying continuous dangerous samples is 0.93. Therefore, 0.08 is used as the valid fluctuation judgment value.

[0259] The processing unit divides the number of valid fluctuations within 2 hours by 2 to obtain the fluctuation frequency, which is measured in times per hour.

[0260] Specifically, the attenuation amplitude is calculated based on the signal strength and fluctuation frequency.

[0261] The processing unit divides a preset time window into a front segment and a back segment. The front segment includes the first 12 sampling points, and the back segment includes the last 12 sampling points. The average signal strength of the front segment is obtained by adding the first 12 signal strengths and dividing by 12. The average signal strength of the back segment is obtained by adding the last 12 signal strengths and dividing by 12. The processing unit subtracts the average signal strength of the back segment from the average signal strength of the front segment to obtain the intensity decrease value. When the intensity decrease value is less than 0, it indicates that the signal strength of the back segment has not attenuated, and the processing unit corrects the intensity decrease value to 0.

[0262] Subsequently, the processing unit includes the intensity decrease value at 0.70 and the fluctuation frequency normalization result at 0.30, and adds the two included results to obtain the attenuation amplitude; the fluctuation frequency normalization result is obtained by dividing the current fluctuation frequency by the maximum fluctuation frequency in the training samples; when the calculation result is greater than 1, it is corrected to 1.

[0263] Furthermore, when determining the duration of the second danger signal by the attenuation amplitude, the processing unit simultaneously reads the average signal strength of the window.

[0264] Specifically, when the attenuation amplitude is less than or equal to 0.20 and the average signal strength of the window is greater than 0.60, the continuous state is determined as a continuous state; when the attenuation amplitude is greater than 0.20 and less than or equal to 0.40, the continuous state is determined as an attenuation observation state; when the attenuation amplitude is greater than 0.40, the continuous state is determined as a rapid attenuation state. The attenuation amplitude thresholds of 0.20 and 0.40 are determined by verification using historical labeled samples: when 0.20 is used as the continuous state determination value, the accuracy rate for identifying continuous dangerous samples is 0.92; when 0.40 is used as the rapid attenuation state determination value, the accuracy rate for identifying instantaneous disturbance samples is 0.91. Therefore, 0.20 and 0.40 are used as the continuous state division values.

[0265] In this embodiment, when the processing unit determines whether the continuous state meets the preset confirmation conditions, it performs five judgments in sequence. First, the number of sampling points with a signal strength greater than 0.60 among the 24 sampling points reaches 20; second, the average signal strength of the window is greater than 0.60; third, the attenuation amplitude is less than or equal to 0.20; fourth, the fluctuation frequency is less than or equal to 3 times per hour; fifth, the average number of abnormal measurement points in the later stage is greater than or equal to the average number of abnormal measurement points in the earlier stage.

[0266] The average number of abnormal measurement points in the latter stage is obtained by adding the number of abnormal measurement points in the last 12 sampling points and dividing by 12. The average number of abnormal measurement points in the former stage is calculated in the same way. When all five conditions are met, the continuous state is determined to meet the preset confirmation conditions. The signal strength threshold of 0.60 is consistent with the danger signal determination threshold in step S5. The number of confirmed sampling points of 20 is determined by verifying 18, 20 and 22 candidate values. Among them, 20 corresponds to a continuous danger signal recognition accuracy of 0.93 and a false recognition rate of 0.06. The fluctuation frequency threshold of 3 times per hour is determined by historical instantaneous disturbance samples. When the threshold is 3 times per hour, the instantaneous disturbance sample recognition accuracy is 0.91.

[0267] Furthermore, when the continuous state meets the preset confirmation conditions, the processing unit retains the second danger signal and generates a continuous confirmation record.

[0268] Specifically, the continuous confirmation record includes position number, measurement point number, window start and end time, second critical value sequence, signal strength sequence, window average signal strength, fluctuation frequency, attenuation amplitude, duration and confirmation result; the retained second danger signal enters S8 as the data source for updating the warning standard.

[0269] Furthermore, when the continuous state does not meet the preset confirmation conditions, the processing unit filters or downgrades the second danger signal.

[0270] Specifically, when the average signal strength of the window is less than or equal to 0.50, or the attenuation amplitude is greater than 0.40, or the number of sampling points with a signal strength greater than 0.60 out of 24 sampling points is less than 12, the processing unit filters the second danger signal, deletes its warning trigger mark, and saves the corresponding record as a historical observation record; when the average signal strength of the window is greater than 0.50 and less than or equal to 0.60, or the attenuation amplitude is greater than 0.20 and less than or equal to 0.40, or the number of sampling points with a signal strength greater than 0.60 out of 24 sampling points is 12 to 19, the processing unit downgrades the second danger signal to an observation signal and generates a review prompt.

[0271] The above filtering and degradation boundaries were obtained through verification using labeled samples: an average window signal strength of 0.50 corresponds to an instantaneous disturbance filtering accuracy of 0.90, and an average window signal strength of 0.60 corresponds to a continuous danger signal confirmation accuracy of 0.93. Therefore, 0.50 and 0.60 are used as the filtering and degradation judgment boundaries.

[0272] Example 9:

[0273] S8 includes obtaining the retained second danger signal, extracting features of the second danger signal to obtain an initial warning standard; calculating the deviation between the first critical value and the features within the initial warning standard, and determining the second critical value; replacing the first critical value with the second critical value, obtaining the initial warning standard based on the retained second danger signal, and obtaining the updated warning standard by replacing the first critical value in the initial warning standard with the second critical value.

[0274] In this embodiment, the processing unit only receives the second danger signal that has been confirmed and retained by S7.

[0275] Specifically, the recorded content of the second danger signal includes the warehouse number, measuring point number, cargo type, window start time, window end time, first critical value, second critical value sequence, signal strength sequence, fluctuation frequency, attenuation amplitude, abnormal measuring point number sequence, and continuous status. The processing unit first verifies the confirmation result of the second danger signal. If the confirmation result is to retain it, the record is written into the early warning standard update queue. If the confirmation result is to filter or downgrade, the critical value replacement is not performed.

[0276] Furthermore, the processing unit extracts features from the retained second hazard signal for updating the warning criteria.

[0277] Specifically, the features include the average value of the second critical value, the peak value of the second critical value, the average signal strength of the window, the average signal strength of the latter part, the attenuation amplitude, the fluctuation frequency, the proportion of abnormal measurement points, and the duration. The average value of the second critical value is obtained by adding the second critical values ​​of 24 sampling points within a preset time window and dividing by 24. The peak value of the second critical value is obtained by reading the maximum value among the 24 second critical values. The average signal strength of the window is obtained by adding the signal strengths of 24 sampling points and dividing by 24. The average signal strength of the latter part is obtained by adding the signal strengths of the last 12 sampling points and dividing by 12. The proportion of abnormal measurement points is obtained by dividing the number of measurement points that generate the second danger signal within the window by the total number of measurement points in the warehouse. The duration is obtained by multiplying the number of sampling points that meet the continuous confirmation condition by 5 minutes.

[0278] In this embodiment, the initial warning criteria are invoked according to the type of goods, warehouse number, and monitoring point number.

[0279] Specifically, the processing unit reads the initial warning standard formed and saved in S5. The initial warning standard includes an observation interval, a warning interval, and a danger interval. The observation interval corresponds to a first critical value greater than 0.60 and less than or equal to 0.75, the warning interval corresponds to a first critical value greater than 0.75 and less than or equal to 0.90, and the danger interval corresponds to a first critical value greater than 0.90 and less than or equal to 1. The processing unit compares the average value of the second critical value in the retained second danger signal with the above three intervals one by one, and determines the interval where the average value of the second critical value is located as the interval to be updated.

[0280] Furthermore, once the interval to be updated is determined, the processing unit reads the original first critical value within that interval.

[0281] Specifically, the first critical value is the critical value saved when S5 determines the first danger signal, and it corresponds one-to-one with the cargo type, warehouse number, measuring point number, and warning interval. In order to ensure that the continuous confirmation result of the second danger signal can enter the warning standard, the processing unit first calculates the feature correction value. During the calculation, the average value of the second critical value is calculated as 0.40, the average signal strength of the window is calculated as 0.30, the proportion of abnormal measuring points is calculated as 0.20, and the result obtained by subtracting the attenuation amplitude from 1 is calculated as 0.10. Then, the results of the four calculations are added together to obtain the feature correction value. The closer the attenuation amplitude is to 0, the more stable the second danger signal is within the preset time window. Therefore, the result obtained by subtracting the attenuation amplitude from 1 is used for correction.

[0282] Specifically, the weights of the feature correction values ​​are determined using historical labeled samples.

[0283] The processing unit calls upon the second hazard signal samples that have been continuously confirmed and retained within the 12 months prior to the current model training date, and uses manual inspection records, ventilation records, dehumidification records, pile turning records, and emergency work orders as verification criteria. Three sets of candidate weights are used for verification. When the average weight of the second critical value is 0.40, the average signal strength of the window is 0.30, the proportion of abnormal measurement points is 0.20, and the weight of the signal retention item is 0.10, the accuracy rate of the early warning standard update is 0.92 and the false recognition rate is 0.06. Therefore, this set of weights is adopted.

[0284] Specifically, the processing unit calculates the deviation between the first critical value and the feature within the initial warning standard.

[0285] The calculation process is as follows: subtract the original first critical value within the interval to be updated from the feature correction value to obtain the critical deviation; when the critical deviation is greater than 0, it indicates that the retained second danger signal has a basis for upward adjustment relative to the original first critical value; when the critical deviation is less than or equal to 0, it indicates that the second danger signal does not trigger the critical value replacement; the processing unit sets the critical deviation trigger value to 0.05, and only when the critical deviation reaches 0.05 will it enter the second boundary value determination process; this trigger value is obtained by verifying three candidate values: 0.03, 0.05, and 0.08; when the candidate value is 0.03, the false recognition rate is 0.10; when the candidate value is 0.05, the early warning standard update accuracy rate is 0.92, and the false recognition rate is 0.06; when the candidate value is 0.08, the missed update rate is 0.09, therefore 0.05 is adopted as the critical deviation trigger value.

[0286] Specifically, the second critical value is determined based on the original first critical value and the critical deviation.

[0287] Specifically, when the critical deviation reaches 0.05, the processing unit multiplies the critical deviation by 0.50 to obtain the critical correction amount, and then adds the critical correction amount to the original first critical value to obtain the second critical value used for replacement.

[0288] Specifically, the critical correction coefficient of 0.50 was determined through verification using three candidate values: 0.30, 0.50, and 0.70. When the candidate value is 0.30, the accuracy rate for identifying subsequent danger signals after the standard update is 0.89; when the candidate value is 0.50, the accuracy rate is 0.92 and the false recognition rate is 0.06; when the candidate value is 0.70, the false recognition rate is 0.09. Therefore, 0.50 was adopted as the critical correction coefficient.

[0289] Furthermore, to prevent the second critical value from exceeding its range, the processing unit limits the calculation results to a specific range.

[0290] Specifically, when the interval to be updated is the observation interval, the second critical value is corrected to 0.60 if it is less than 0.60, and to 0.75 if it is greater than 0.75; when the interval to be updated is the warning interval, the second critical value is corrected to 0.75 if it is less than 0.75, and to 0.90 if it is greater than 0.90; when the interval to be updated is the danger interval, the second critical value is corrected to 0.90 if it is less than 0.90, and to 1 if it is greater than 1; the second critical value after being defined is written into the warning standard as a replacement value.

[0291] During the replacement process, the processing unit locates the record in the initial warning standard that corresponds to the type of goods, warehouse number, measurement point number, and interval to be updated, and replaces the first critical value field in that record with the second critical value.

[0292] Meanwhile, the processing unit saves the first critical value, feature correction value, critical deviation, critical correction amount, second critical value and update time before the update; if the same type of goods, the same warehouse location, and the same measuring point have more than two retained second danger signals within 24 hours, the processing unit calculates the corresponding second critical value for each time, adds up all the second critical values ​​and divides them by the number of times, and uses the average result as the final replacement value.

[0293] Furthermore, after the replacement is completed, the processing unit generates updated warning standards. The updated warning standards include cargo type, warehouse location number, monitoring point number, observation interval, warning interval, danger interval, update threshold value for each interval, applicable start time, and version number. The version number increments sequentially according to the update time.

[0294] Example 10:

[0295] S9 includes acquiring warehousing environment data, determining a second critical value for the target dry bulk cargo based on the warehousing environment data; calculating the degree of deviation between the second critical value and the updated warning standard, wherein the updated warning standard is obtained by fitting historical warning records; determining whether the degree of deviation is greater than a preset deviation threshold; if the degree of deviation is greater than the preset deviation threshold, determining the warehousing abnormality level of the target dry bulk cargo based on the degree of deviation, and outputting the graded warning information corresponding to the warehousing abnormality level.

[0296] In this embodiment, when the processing unit executes S9, it first obtains the storage environment data of the warehouse where the target dry bulk cargo is located.

[0297] Specifically, the storage environment data includes the following: warehouse temperature, relative humidity, dew point temperature, ventilation speed, temperature difference between inside and outside the warehouse, air pressure, surface temperature of goods, moisture content of goods, number of abnormal measuring points, and operating status of ventilation equipment. The sampling cycle is 5 minutes. The processing unit reads the current sampling time and the sampling records from the previous 2 hours according to the warehouse location number and measuring point number, forming a data segment of 24 sampling points. If a single sampling point is missing, the processing unit reads the previous and next valid values ​​of that sampling point, adds the two valid values ​​together, and divides by 2 to obtain the supplementary value for that sampling point. If three consecutive sampling points are missing, the processing unit removes the data of that measuring point within that time segment and calls the data of other valid measuring points in the same warehouse location to participate in the calculation.

[0298] In this embodiment, the processing unit determines the second critical value of the target dry bulk cargo based on the storage environment data.

[0299] Specifically, the following parameters are calculated: temperature change rate, humidity change rate, dew point difference, moisture content change, and ventilation speed change. The temperature change rate is obtained by subtracting the temperature inside the warehouse from the temperature at the previous sampling time, and then dividing by 5 minutes. The humidity change rate is obtained by subtracting the relative humidity at the previous sampling time from the current relative humidity, and then dividing by 5 minutes. The dew point difference is obtained by subtracting the surface temperature of the goods from the dew point temperature. The moisture content change is obtained by subtracting the moisture content at the previous sampling time from the current moisture content. The ventilation speed change is obtained by subtracting the ventilation speed at the previous sampling time from the current ventilation speed.

[0300] The processing unit correlates the above calculation results with the core meteorological data, meteorological correction value and signal strength in the second danger signal retained in S7, and generates risk correction data for the current sampling time.

[0301] Furthermore, the calculation of the second critical value continues the critical correction logic formed in S6 and S8.

[0302] Specifically, the processing unit reads the first critical value corresponding to the current cargo status sequence and calculates the environmental correction value based on the storage environment data. The calculation process for the environmental correction value is as follows: the normalized result of the humidity change rate is recorded as 0.25, the normalized result of the temperature change rate is recorded as 0.20, the dew point difference risk item is recorded as 0.20, the normalized result of the moisture content change is recorded as 0.20, and the normalized result of the ventilation speed change is recorded as 0.15. Then, the results of the five items are added together. The dew point difference risk item is determined as follows: when the result of the dew point temperature minus the cargo surface temperature reaches 0℃ or is greater than 0℃, the dew point difference risk item is recorded as 1; when it is less than 0℃, the dew point difference risk item is recorded as 0. During normalization, the current index value is subtracted from the minimum value of the corresponding index in the training sample, and then divided by the difference between the maximum and minimum values ​​of the corresponding index in the training sample.

[0303] Furthermore, after obtaining the environmental correction value, the processing unit calculates the first critical value as 0.70, calculates the environmental correction value as 0.30, and adds the two calculation results to obtain the second critical value of the target dry bulk cargo.

[0304] If the obtained second critical value is greater than 1, it is corrected to 1; if the obtained second critical value is less than 0, it is corrected to 0. The ratios of 0.70 and 0.30 are determined by historical sample verification. The processing unit uses the labeled warning records within 12 months before the current model training date as the verification object, and verifies them using three sets of candidate ratios: 0.80 and 0.20, 0.70 and 0.30, and 0.60 and 0.40, respectively. When the first critical value accounts for 0.70 and the environmental correction value accounts for 0.30, the anomaly level identification accuracy is 0.93 and the false identification rate is 0.06. Therefore, this ratio is adopted.

[0305] Furthermore, the updated warning criteria were obtained by fitting historical warning records.

[0306] Specifically, the processing unit reads historical early warning records corresponding to the same cargo type, storage type, and disposal rules as the target dry bulk cargo within 12 months prior to the current model training date. Each historical early warning record includes the second critical value, signal strength, deviation degree, manual inspection conclusion, disposal measures, emergency work order status, and final anomaly level. The processing unit first removes records whose confirmation result is "filtered," and then retains records whose confirmation result is "retain" or "downgrade" as fitting samples. During fitting, the second critical value and signal strength are used as independent variables, and the level value corresponding to the final anomaly level is used as the dependent variable, where normal state is recorded as 0, level 1 early warning as 1, level 2 early warning as 2, and level 3 early warning as 3.

[0307] As a preferred embodiment, the processing unit uses a piecewise fitting method to generate updated warning criteria.

[0308] Specifically, historical early warning records are first divided into Level 1, Level 2, and Level 3 early warning samples according to the final anomaly level. Then, the average value and boundary value of the second critical value in each level sample are calculated. The average value is obtained by adding all the second critical values ​​in that level and dividing by the number of samples. The boundary value is obtained by taking the median value of the average values ​​of adjacent level samples. Specifically, the average value of the second critical values ​​of the Level 1 early warning samples and the average value of the second critical values ​​of the Level 2 early warning samples are added and divided by 2 to obtain the level boundary between Level 1 and Level 2 early warnings. The average value of the second critical values ​​of the Level 2 early warning samples and the average value of the second critical values ​​of the Level 3 early warning samples are added and divided by 2 to obtain the level boundary between Level 2 and Level 3 early warnings. The processing unit compares the fitted level boundary with the updated critical value formed by S8 and takes the average of the two as the final boundary value in the updated early warning standard.

[0309] Furthermore, when calculating the degree of deviation between the second critical value and the updated warning standard, the processing unit first determines the warning interval in which the second critical value is located.

[0310] The updated early warning standards include observation intervals, Level 1 early warning intervals, Level 2 early warning intervals, and Level 3 early warning intervals. The observation interval corresponds to a second critical value greater than 0.50 and less than or equal to the Level 1 early warning boundary. The Level 1 early warning interval corresponds to a second critical value greater than the Level 1 early warning boundary and less than or equal to the Level 2 early warning boundary. The Level 2 early warning interval corresponds to a second critical value greater than the Level 2 early warning boundary and less than or equal to the Level 3 early warning boundary. The Level 3 early warning interval corresponds to a second critical value greater than the Level 3 early warning boundary and less than or equal to 1. The processing unit subtracts the lower limit of the interval from the second critical value and divides it by the difference between the upper and lower limits of the interval to obtain the degree of deviation of the second critical value within that interval. If the calculation result is greater than 1, it is corrected to 1; if the calculation result is less than 0, it is corrected to 0.

[0311] Specifically, the preset deviation threshold is set to 0.20.

[0312] The threshold was obtained by verifying historical early warning records. The processing unit used 0.15, 0.20, and 0.25 as candidate deviation thresholds, respectively, and compared the calculated early warning level with the manual inspection conclusions, warehouse disposal records, and emergency work order records. When the candidate deviation threshold was 0.15, the false recognition rate was 0.11. When the candidate deviation threshold was 0.20, the accuracy rate of warehouse abnormal status level identification was 0.93, and the false recognition rate was 0.06. When the candidate deviation threshold was 0.25, the missed identification rate was 0.09. Therefore, the processing unit used 0.20 as the deviation threshold.

[0313] Furthermore, when the processing unit determines whether the deviation degree is greater than the preset deviation threshold, it compares the deviation degree with 0.20. If the deviation degree is less than or equal to 0.20, the processing unit does not raise the warning level and saves the current record as a tracking record. If the deviation degree is greater than 0.20, the processing unit determines the storage abnormality status level of the target dry bulk cargo based on the deviation degree. If the deviation degree is greater than 0.20 and less than or equal to 0.50, it is determined to be a Level 1 warning. If the deviation degree is greater than 0.50 and less than or equal to 0.80, it is determined to be a Level 2 warning. If the deviation degree is greater than 0.80 and less than or equal to 1, it is determined to be a Level 3 warning.

[0314] Among them, the above-mentioned level boundaries of 0.50 and 0.80 were determined by fitting and verifying historical early warning records. 0.50 corresponds to an accuracy rate of 0.92 for identifying warehousing and disposal records, and 0.80 corresponds to an accuracy rate of 0.91 for identifying emergency work order records.

[0315] Furthermore, the processing unit outputs tiered early warning information based on the level of abnormal warehouse conditions.

[0316] Specifically, the tiered early warning information includes cargo type, warehouse location number, monitoring point number, anomaly level, second critical value, deviation degree, trigger time, anomaly type, and handling measures. The handling measures for a Level 1 early warning are to generate an on-site verification task and continue tracking the second critical value in the next sampling cycle. The handling measures for a Level 2 early warning are to activate ventilation equipment, dehumidification equipment, or perform localized turning operations and generate a handling record. The handling measures for a Level 3 early warning are to generate an emergency work order, isolate the corresponding warehouse area, and notify on-duty personnel for verification. Output methods include display on the warehouse management terminal, audio-visual prompts, mobile terminal messages, and early warning record entry.

[0317] Example 11:

[0318] Figure 2 This is a structural block diagram of the local terminal of an exemplary electronic device (machine) of the present invention; as shown... Figure 2 As shown, the electronic device of the present invention includes a processor 11, a memory 12, a storage space 13 for storing program code, and program code 14 for executing the method steps according to the present invention. The program code 14 for executing the method steps according to the present invention is used to execute the above-described control logic.

[0319] Figure 3 This is a structural block diagram of the network end of an exemplary electronic device of the present invention; as shown below. Figure 3 As shown, the present invention also provides an electronic device (machine), which may include at least one processor 210, at least one memory 230 communicatively connected to the processor, and a communication bus 240 and a communication interface 220 connecting different system components (including the memory 230 and the processor 210). The processor 210, the memory 230 and the communication interface 220 are connected through the communication bus 240 and communicate with each other. The communication interface 220 is used for data interaction with external devices. The memory 230 stores a machine-executable program that can be executed by the processor, and the processor 210 can execute the above-mentioned control logic by calling the machine-executable program.

[0320] Communication bus 240 represents one or more of several bus architectures, including a memory bus or memory controller, peripheral bus, graphics acceleration port, processor, or local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MAC) bus, Enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnection (PCI) bus.

[0321] Electronic devices typically include a variety of computer system readable media, which can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, and removable and non-removable media.

[0322] Memory 230 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The electronic device may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 230 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the control logic described above.

[0323] A program / utility having a set (at least one) of program modules can be stored in memory 230. Such program modules include, but are not limited to, an operating system, one or more applications, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment.

[0324] Machine-executable programs for performing this invention can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages ​​such as Java, C++, and Python, and may also include specialized engineering languages ​​such as R. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0325] The present invention also discloses a storage medium on which a machine-executable program as described above is stored.

[0326] The aforementioned storage medium may be any combination of one or more computer-readable media. Computer-readable media may be, for example, computer-readable signal media or computer-readable storage media. Computer-readable storage media include, but are not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium may be, for example, any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0327] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0328] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0329] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification and low false alarm tiered early warning of abnormal states in dry bulk cargo storage, characterized in that, include: S1. Obtain the type and physicochemical properties of the target dry bulk cargo, obtain the climate conditions in the storage space, process the type and physicochemical properties of the cargo and the climate conditions, and obtain the interaction characteristics. The interaction characteristics include the interaction characteristics between the moisture absorption rate and the change in relative humidity, the interaction characteristics between the moisture content and the rate of temperature change, the interaction characteristics between porosity and the ventilation speed, and the interaction characteristics between the moisture absorption and heat release coefficient and the dew point difference. S2. Based on the interaction characteristics, obtain environmental fluctuation data within the storage space, perform multi-scale decomposition on the environmental fluctuation data, and obtain the cargo state sequence. S3. Obtain historical experience data and determine the initial safety boundary by comparing the cargo status sequence with historical experience data; S4. If there is a cargo state sequence that deviates from historical experience data within the initial safety boundary, extract the risk evolution features corresponding to the cargo state sequence that deviates from historical experience data. S5. Obtain the initial warning standard. By calculating the deviation value between the risk evolution characteristics and the initial warning standard, a first critical value is obtained. If the first critical value exceeds the preset threshold, the cargo status sequence corresponding to the first critical value is determined to be the first danger signal. The initial warning standard is obtained by comparing the risk evolution characteristics with the historical safety baseline. S6. For the first danger signal, obtain real-time climate conditions, and obtain a second critical value by fusing the first danger signal and the real-time climate conditions. If the second critical value is greater than the first critical value, then extract the second danger signal corresponding to the second critical value. S7. Confirm the second danger signal. If the second danger signal continuously meets the preset confirmation conditions within the preset time window, the second danger signal is retained. If the second danger signal does not continuously meet the preset confirmation conditions, the second danger signal is filtered or downgraded. S8. Based on the retained second danger signal, obtain the initial warning standard. By replacing the first critical value in the initial warning standard with the second critical value, obtain the updated warning standard. Read the initial warning standard formed and saved in S5. The initial warning standard includes an observation interval, a warning interval, and a danger interval. Compare the average value of the second critical value in the retained second danger signal with the above three intervals one by one. Determine the interval where the average value of the second critical value is located as the interval to be updated. After the interval to be updated is determined, read the original first critical value in the interval. S1 includes: Acquire the cargo type and initial physicochemical properties data of the target dry bulk cargo, clean the initial physicochemical properties data to obtain standard physicochemical properties characteristics, and acquire the initial climate conditions data within the storage space; A feature extraction algorithm is used to process standard physicochemical properties and initial climate conditions data to obtain a multidimensional environmental cargo feature matrix. If the feature dimension of the multidimensional environmental cargo feature matrix exceeds the preset dimension threshold, the multidimensional environmental cargo feature matrix is ​​reduced in dimension by principal component analysis algorithm to obtain the core feature vector. The input dataset is constructed based on the core feature vectors. The random forest algorithm is used to perform node splitting calculations on the input dataset to obtain the interaction characteristics of cargo types, physicochemical properties, and climatic conditions of the target dry bulk cargo. S2 includes: Obtain an environmental fluctuation data matrix based on interaction characteristics; The environmental fluctuation data matrix is ​​decomposed to obtain multi-scale environmental decomposition components. The multi-scale environment is decomposed and its components are reduced in dimensionality to obtain the core environmental feature vector, and the hygroscopic and exothermic physical properties are obtained to construct boundary constraint conditions. An input tensor is constructed based on the core environmental feature vector and boundary constraints. The input tensor is then processed through a long short-term memory network to obtain the cargo state sequence.

2. The method for intelligent identification and low false alarm graded early warning of abnormal states in dry bulk cargo storage according to claim 1, characterized in that: S3 includes: Obtain the cargo status sequence and calculate the trajectory similarity between the cargo status sequence and historical experience data; Based on trajectory similarity, historical experience data of the target is determined, and the state fluctuation range is obtained by comparing the cargo state sequence with the historical experience data of the target. If the empirical matching degree corresponding to the state fluctuation amplitude is lower than the threshold, an abnormal state node is generated. Calculate the boundary deviation value based on the abnormal state node, and determine the initial safety boundary based on the boundary deviation value.

3. The method for intelligent identification and low false alarm graded early warning of abnormal states in dry bulk cargo storage according to claim 1, characterized in that: S4 includes: The difference between the cargo state sequence that deviates from historical experience data and the baseline state is calculated to obtain the state offset vector; The temporal evolution gradient is obtained from the state offset vector, and the temporal evolution gradient is fused with environmental factors to determine the risk evolution rate. If the rate of risk evolution is greater than the threshold, then a trend matrix is ​​constructed based on the rate of risk evolution. The trend matrix is ​​dimensionality reduced to obtain the risk evolution characteristics of the cargo state sequence that deviates from historical experience data.

4. The method for intelligent identification and low false alarm graded early warning of abnormal states in dry bulk cargo storage according to claim 1, characterized in that: S5 includes: The initial early warning criteria are obtained by comparing risk evolution characteristics with historical safety baselines. Calculate the projection distance of the risk evolution characteristics within the interval corresponding to the initial warning standard to obtain the deviation value; The first critical value is obtained based on the deviation value; Determine whether the first critical value exceeds the preset threshold; If the first critical value exceeds the preset threshold, the cargo state sequence corresponding to the first critical value is determined to be the first danger signal.

5. The method for intelligent identification and low false alarm graded early warning of abnormal states in dry bulk cargo storage according to claim 1, characterized in that: S6 includes: Obtain the initial risk level corresponding to the first danger signal, and obtain real-time climate conditions based on the initial risk level; Real-time climate conditions are dimensionality-reduced to obtain core meteorological data, and fusion weights are determined based on the core meteorological data. The first danger signal and core meteorological data are fused by fusion weighting to obtain the second critical value; Determine whether the second critical value is greater than the first critical value; If the second critical value is greater than the first critical value, then the second danger signal corresponding to the second critical value is extracted.

6. The method for intelligent identification and low false alarm graded early warning of abnormal states in dry bulk cargo storage according to claim 1, characterized in that: S7 includes: The second danger signal is acquired and its feature sequence is extracted. The feature sequence is processed within a preset time window to obtain the signal strength and fluctuation frequency. The attenuation amplitude is calculated based on the signal strength and fluctuation frequency, and the duration of the second danger signal is determined by the attenuation amplitude. Determine whether the persistent state meets the preset confirmation conditions; If the persistent state meets the preset confirmation conditions, the second danger signal is retained; If the persistent state does not meet the preset confirmation conditions, the second danger signal will be filtered or downgraded.

7. The method for intelligent identification and low false alarm graded early warning of abnormal states in dry bulk cargo storage according to claim 1, characterized in that: S8 includes: Obtain the retained second danger signal, and extract the features of the second danger signal to obtain the initial warning standard; Calculate the deviation between the first critical value and the characteristic within the initial warning standard, and determine the second critical value; The first critical value is replaced by the second critical value. The initial warning standard is obtained based on the retained second danger signal. The updated warning standard is obtained by replacing the first critical value in the initial warning standard with the second critical value.

8. The method for intelligent identification and low false alarm graded early warning of abnormal states in dry bulk cargo storage according to claim 1, characterized in that, It also includes S9, which determines the storage anomaly level of the target dry bulk cargo based on the deviation between the second critical value and the updated warning standard, and outputs the graded warning information corresponding to the storage anomaly level, specifically including: Acquire warehousing environment data, and determine the second critical value of the target dry bulk cargo based on the warehousing environment data; Calculate the degree of deviation between the second critical value and the updated warning standard, which is obtained by fitting historical warning records; Determine whether the degree of deviation is greater than the preset deviation threshold; If the deviation exceeds the preset deviation threshold, the storage anomaly level of the target dry bulk cargo is determined based on the deviation level, and a graded early warning message corresponding to the storage anomaly level is output.

Images